When Robots Kill

When Robots Kill
Gabriel Halle v y
Northeaster n University Press . Boston

Northeastern University Press
An imprint of University Press of New England
www.upne.com
© 2013 Northeastern University
All rights reserved
Manufactured in the United States of America
Designed by Vicki Kuskowski
Typeset in Melior by Copperline Book Services, Inc.
University Press of New England is a member of the Green Press

Initiative. The paper used in this book meets their minimum
requirement for recycled paper.
For permission to reproduce any of the material in this book,

contact Permissions, University Press of New England,
One Court Street, Suite 250, Lebanon NH 03766;
or visit www.upne.com
Library of Congress Cataloging-in-Publication Data

Halevi, Gavri’el.
When robots kill : artificial intelligence under criminal law /
Gabriel Hallevy.
pages cm.
Includes bibliographical references and index.
ISBN 978–1–55553–801–9 (cloth) —
ISBN 978–1–55553–805–7 (pbk.) —
ISBN 978–1–55553–806–4 (ebook) (print)
1. Criminal liability. 2. Artificial intelligence—Law and
legislation—Criminal provisions. I. Title.
K5070.H35 2013
345'.0252—dc23 2012033518
5 4 3 2 1
to my dear f amil y
And God said, let us make man in our image, after our
likeness: and let them have dominion over the fish of the sea,
and over the fowl of the air, and over the cattle,
and over all the earth, and over every creeping thing that
creepeth upon the earth. So God created man in
his own image, in the image of God created he him;
male and female created he them.
—Ge nesis 1:26–27
Con te nts
List of Tables . xiii

Preface . xv
1
T he E mer ge nce of M achi na S a pie ns C rimi nalis . 1
1.1. The Endless Quest for Machina Sapiens . 1
1.1.1. History and Prehistory of Artificial Intelligence . 1
1.1.2. Defining Artificial Intelligence, and the Endlessness
of the Quest for Machina Sapiens . 4
1.1.3. Turning Disadvantages into Advantages:
Industrial and Private Use of Artificial Intelligence . 12
1.2. On Evolution and Devolution:
Machina Sapiens Criminalis . 14
1.2.1. Serve and Protect: Human Fears of
Human-Robot Coexistence . 14
1.2.2. Evolution and Devolution:
Machina Sapiens Criminalis as a By-product . 17
1.2.3. Inapplicability of the Zoological
Legal Model to ai Technology . 22
1.3. The Modern Offender . 25
1.3.1. Basic Requirements for a Given Offense . 26
1.3.2. Requirements of a Given Offense:
Modern Criminal Liability as a Matrix of Minimalism . 29
1.3.3. The Case of Nonhuman Corporations (Round 1) . 34
2
a i C rimi nal L ia bility for In te ntio nal O ffe nses . 38
2.1. The Factual Element Requirement . 38
2.1.1. Conduct . 39
ix
con te nts . x
2.1.2. Circumstances . 42
2.1.3. Results and Causal Connection . 43
2.2. The Mental Element Requirement . 44
2.2.1. Structure of the Mens Rea Requirement . 45
2.2.2. Fulfillment of the Cognitive Aspect . 49
2.2.3. Fulfillment of the Volitive Aspect . 56
2.3. Criminally Liable Entities for Intentional ai Offenses . 64
2.3.1. ai Entity Liability . 64
2.3.2. Human Liability: Perpetration-through-Another . 69
2.3.3. Joint Human and ai Entity Liability:
Probable Consequence Liability . 75
Closing the Opening Example: Intentional Killing Robot . 82
3
a i C rimi nal L ia bility for Ne gli ge nce O ffe nses . 84

3.2.1. Structure of Negligence Requirement . 85
3.2.2. Fulfillment of the Negligence Requirement:
Is Objectivity Subjective? . 90
3.3. Criminally Liable Entities for ai Negligence Offenses . 95
3.3.2. Human Liability: Perpetration-through-Another
and Semi-Innocent Agents . 97
Closing the Opening Example: Negligent Killing Robot . 101
4
a i C rimi nal L ia bility for S trict L ia bility O ffe nses . 104

4.2.1. Structure of Strict Liability Requirement . 105
4.2.2. Fulfillment of Strict Liability:
Making the Factual Become Mental . 109
con te nts . xi
4.3. Criminally Liable Entities for

ai Strict Liability Offenses . 112
4.3.2. Human Liability: Perpetration-through-Another . 114
Closing the Opening Example:
Strict Liable Killing Robot . 117
5
App lica bility of Ge neral D efe nses to
a i C rimi nal L ia bility . 120
5.1. The Function of General Defenses in Criminal Law . 121

5.2. Exemptions . 123
5.2.1. Infancy . 124
5.2.2. Loss of Self-Control . 126
5.2.3. Insanity . 128
5.2.4. Intoxication . 130
5.2.5. Factual Mistake . 133
5.2.6. Legal Mistake . 136
5.2.7. Substantive Immunity . 138
5.3. Justifications . 140
5.3.1. Self-Defense . 140
5.3.2. Necessity . 144
5.3.3. Duress . 147
5.3.4. Superior Orders . 149
5.3.5. De Minimis . 152
Robot Killing in Self-Defense . 154
6
S e nte nci ng a i . 156
6.1. Conceptual Applicability of Criminal

Punishment of ai Entities . 157
6.1.1. ai Purpose of Sentencing:
Combining Rehabilitation and Incapacitation . 157
con te nts . xii
6.1.2. The Legal Technique of Conversion:

The Case of Corporations (Round 2) . 162
6.2. Applicability of Punishments to ai Entities . 165
6.2.1. Capital Punishment . 000
6.2.2. Imprisonment and Suspended Imprisonment . 167
6.2.3. Probation . 169
6.2.4. Public Service . 171
6.2.5. Fine . 173
Sentencing the Killing Robot . 175
Conclusions . 177
Notes . 179
Selected Bibliography . 213
Index . 241
t ables
1.1. Target Questions of the Factual Element Components . 31

1.2. Schematic Comparison of the Requirements of
Mental Element Forms . 33
1.3. Schematic Comparison of the Requirement
for Satisfying the Mental Element Forms . 33
2.1. General Structure of Mens Rea . 49
3.1. General Structure of Negligence . 89
4.1. General Structure of Strict Liability . 109
4.2. The Legal State Reflected in the Criminal Mental State of
the Other Person in Perpetration-through-Another . 115
xiii
pref ace
In 1981, a thirty-seven-year-old Japanese employee in a

motorcycle factory was killed by an artificial intelligence
robot working near him. The robot erroneously identified
the employee as a threat to its mission, and calculated that
the most efficient way to eliminate the threat was to push the
worker into an adjacent machine. Usingits very powerful
hydraulic arm, the robot smashed the surprised worker into
the operating machine, killing him instantly, after which it
resumed its duties without further interference. This is not
science fiction, and the legal question is this: Who is to be
held criminally liable for this homicide?1
In most developed countries, unmanned vehicles, surgical robots, indus-

trial robots, trading algorithms, personal robots, and other artificial intel-
ligence (ai ) entities are in common use. Such use may be personal, medi-
cal, military, commercial, or industrial. The question of criminal liability
arises when the unmanned vehicle is involved in car accidents, the surgi-
cal robot is involved in surgical errors, the trading algorithm is involved
in fraud, and so on. Who is to be held criminally liable for these offenses:
the manufacturer, the programmer, the user, or the ai entity itself?
The technological world is changing rapidly. Robots and computers
are more frequently replacing humans in performing simple activities.
As long as humanity used computers as mere tools, there was no signifi-
cant difference between computers and screwdrivers, cars, or telephones.
But as computers became increasingly sophisticated, we started saying
that they “think” for us. Problems began when computers evolved from
“thinking machines” (devices programmed to perform specific thought
processes, such as computing) into thinking machines without the quota-
tion marks — in other words, artificially intelligent. Artificial intelligence
is the ability of a machine to imitate intelligent behavior.
Artificial intelligence, then, is the simulation of human behavior and
cognitive processes on a computer. As such, it is also the study of the nature
of the entire domain of intelligent minds. Artificial intelligence research
xv
pref ace . xvi
began in the 1940s and early 1950s. Since then, ai entities have become an
integral part of modern human life, functioning in a much more sophisti-
cated way than other common tools. Could these entities become danger-
ous? Indeed, they already are, as the incident I recounted earlier attests.
Artificial intelligence and robotics are making their way into everyday
modern life. And some of the questions I have listed are currently before
the courts or regulators. But there is little comprehensive analysis in the
literature about assessing liability for robots, machines, or software that
exercise varying degrees of autonomy.
The objective of this book is to develop a comprehensive, general, and
legally sophisticated theory of the criminal liability for artificial intelli-
gence and robotics. In addition to the ai entity itself, the theory covers the
manufacturer, the programmer, the user, and all other entities involved.
Identifying and selecting analogies from existing principles of criminal
law, the theory proposes specific ways of thinking through criminal li-
ability for a diverse array of autonomous technologies in a diverse set of
reasonable circumstances.
Chapter 1 contains an investigation of some basic concepts on which
the development of subsequent chapters is built. I will address two main
questions:
1. What can be considered as an artificial intelligence entity?

2. What are the necessary requirements for the imposition of crim-
inal liability?
Because criminal law does not depend on moral accountability, the debate
about the moral accountability of machines is irrelevant to these ques-
tions. Although at times there is an overlap between criminal liability,
certain types of moral accountability, and certain types of ethics, such co-
incidence is not necessary in order to impose criminal liability.
The question of the applicability of criminal law to machines involves
two interrelated, secondary questions:
1. Is criminal liability applicable to machines?

2. Is criminal punishment applicable to machines?
The first question is a core issue of criminal law, and imposition of

criminal liability depends on meeting its requirements. Only if these re-
quirements are met does the question of punishability arise, that is, how
human society can punish machines. The answers proposed in this book
are affirmative for both questions. Chapters 2 through 5 examine the first
question, and chapter 6 examines the second. If the affirmative answers
pref ace . xvii
offered in this book are accepted, then a new social entity has indeed been
created, machina sapiens criminalis. The emergence of machina sapiens
criminalis is examined in chapter 1. Finally, this study answers the ques-
tion of whether artificial intelligence entities could ever be considered
legally, if not morally, liable for the commission of criminal offenses, and
deserving of punishment.
This book focuses on the criminal liability of ai entities and does not
formally venture into the related areas of ethics (including roboethics) and
morality. Nevertheless, the discussion of criminal liability presented here
sets the stage for an in-depth consideration of the ethical issues involved,
which will be necessary given the proliferation of ai entities in everyday
human life. This book constructs a suitable framework for concrete and
functional discussion of these issues.
I wish to thank Dr. Phyllis D. Deutsch of University Press of New En-
gland for her faith in this project from its very beginning, for guiding the
publication of the book from inception to conclusion, and for her useful
comments. I also wish to thank Gabriel Lanyi for his comments, insights,
and disputations. I thank the anonymous readers for their comments. I also
thank Ono Academic College and Northeastern University for their gener-
ous support of this project. Finally, I wish to thank my wife and daughters
for their staunch support all along the way.
—G. H .
When Robots Kill
1
The E mergence of M achina Sapien s C riminali s
1.1. The Endless Quest for Machin a Sapiens
1.1.1. H istory and P rehistory of A rtificial I ntelligence

Since the dawn of humanity, we have sought tools to make life easier. In
the Stone Age, tools were made of stone. As humans discovered the advan-
tages of metals, metal replaced stone. With the expansion of knowledge,
tools proliferated and played increasing roles in daily human life. Tools
were continually challenged by increasingly complicated tasks. When
they failed, more sophisticated tools were invented; if these succeeded,
new challenges were posed for them, in a perpetual cycle that continues to
this day. Mechanical devices have been used to ease daily life since antiq-
uity. Heron of Alexandria, in the first century ad , designed the first steam
engine and a wind-powered organ, among his many other inventions.1
But machines became commonly used in mass production during the first
industrial revolution, in the eighteenth century.
The idea of thinking machines evolved together with the insight into
the human ability to create systematic methods of rational thinking. Des-
cartes initiated the human quest for such methods in 1637,2 although he
did not believe that mechanical devices could achieve reason.3 But Des-
cartes laid the groundwork for the symbol-processing machines of the
modern age. In 1651, Hobbes described reason as symbolic calculation.4
During the seventeenth century, Leibniz held forth the hope of discovering
a universal language of mathematics, science, and metaphysics, charac-
teristica universalis, which would enable replacing thinking with calcula-
tion, and Pascal designed adding and multiplying machines, most likely
the first mechanical computers,5 operated entirely by humans and inca-
pable of thinking. Naturally, these machines were not expected to think
independently.
The modern idea of a thinking machine is generally associated with
Ada Lovelace, the daughter of Lord Byron and patroness of Charles Bab-
1
When R obo ts Kill . 2
bage, inventor and designer of the analytical engine — the first mechanical,

programmable computer, but never completed. In 1843, when she was ex-
posed to Babbage’s work, Lady Lovelace asked Babbage whether his ma-
chine could actually “think.” The notion of mechanical thinking was ex-
tremely odd in those days, but nevertheless, the question itself opened the
mind for considering the feasibility of unnatural, or “artificial” intelligence.
But the idea of artificial intelligence could not be examined in practice
until electricity, and eventually electronic computers, became reality. In
the 1950s, momentous developments in machine-to-machine translation
made it possible for computers to communicate with each other, and fur-
ther developments in human-to-machine translation made it possible for
human operators to communicate with computers, initially in a limited
fashion. In time, computer scientists learned to analyze reasoning and rep-
resent knowledge using symbols to advance the important area of artificial
intelligence known as knowledge representation.6
The ability of electronic computers to store large amounts of informa-
tion and process it at high speeds challenged scientists to build systems
that could exhibit human capabilities. Since the 1950s, human skills and
capabilities have increasingly been captured and carried out by electronic
machines. With the development and popularization of the personal com-
puter, these capabilities have become affordable for many people. Con-
tinually increasing memory capacities, speed, reliability, and robustness
of personal computers are today bringing artificially intelligent tools to
desktops worldwide.
During the rapid advances in the field of computers in the 1950s, ar-
tificial intelligence (ai ) developed as a separate area of research that com-
bined technological investigation with the study of logic and eventually
of cybernetics, an interdisciplinary field dealing with communication be-
tween humans and machines, among other issues. Studies in logic in the
1920s and 1930s helped formalize the methods of reasoning, and pro-
duced a new form of logic known as propositional and predicate calculus,
relying on the works of Church, Gödel, Post, Russell, Tarski, Whitehead,
Kleene, and others.7 Developments in psychology, neurology, statistics,
and mathematics in the 1950s were also incorporated into the growing
research sphere of ai .8
In the late 1950s, several developments took place that signaled the
emergence of ai to the public at large. The most prominent of these was the
development of chess-playing programs and of the General Problem Solver
(cpr ), designed to solve a wide range of problems, from symbolic integra-
tion to word puzzles. This brought the abilities of ai to the attention of an
The Emerge nce of Machin a Sapiens C r iminalis . 3
enthusiastic public, creating at the same time unrealistic expectations and

also the first science fiction novels about robots rebelling against humans
and establishing control over them.9
ai research proceeded in two main directions: (1) building physical
devices combined with digital computers, and (2) developing symbolic
representations. The first effort led to robotics; the second sought to train
perception to classify certain types of patterns as either similar or dis-
tinct. ai research was guided by the assumption that the problem of com-
monsense knowledge is solvable, and that if humans can solve it, so can
machines. The problem was formulated as the ability to glean facts in
situations when not all the information is given, using common sense. For
many years, this assumption prevented progress in theoretical ai and di-
verted ai research toward connectionism (neural network models).10
By the 1970s, the importance of ai became apparent to most of the
world. Governments in developed and developing countries were seek-
ing long-term commitments of resources to fund intensive research pro-
grams in ai .11 Government and private organizations routinely cooperated
in development programs in the areas of robotics, software, and various
computer products. These projects were driven by the realization that it
had become feasible to develop systems that could exhibit such human
abilities as understanding speech and visual scenes, learning and refining
knowledge, and making independent decisions.
Beginning in the 1970s, industry embraced ai technology. Some ai ca-
pabilities, such as natural language processing, have not yet been fully
implemented, but in a growing number of fields, ai technology is being
applied successfully by industry.12 Examples of such areas are biomedical
microscopy, materials analysis, and robotics. Areas in which ai has been
successfully applied have become entirely dependent on this technology
because common human abilities cannot replace the accuracy, speed, and
efficiency of these ai tools.
In the 1980s, ai research made giant gains in the design and develop-
ment of expert systems in the fields of medicine, finance, and anthropol-
ogy. The main challenge of expert systems was to develop suitable knowl-
edge representations in their respective fields. To be readily accessible,
this knowledge needed to be stored in a form that could be retrieved and
displayed automatically by the system interfaces, human or other. Many
expert systems became successful tools, expanded over the years with new
knowledge and improved with better heuristics.13
The next challenge was to enable new technologies to be incorporated
into expert systems shortly after they became available. Development of
expert systems resulted in intensive research in the areas of machine learn-

ing and problem solving, which expanded the use of expert systems into
many new fields, with development efforts extending beyond the confines
of academic research, and ai technology increasingly replacing tradition-
ally human capabilities.14 With the growing need for ai development in in-
dustry, the advantages of ai technology were no longer subjects of debate.
The involvement of industry in ai research grew for three main rea-
sons. First, the achievements of ai , especially in knowledge engineering,
were beyond any doubt. Second, advances in hardware made the develop-
ment of ai tools more feasible and the products more accessible to users.
Third, industry had a growing need for faster and better problem-solving
tools in order to increase productivity. Industry, therefore, adopted the use
of ai technology in many areas, such as factory automation, programming,
office automation, and personal computing.15
The combination of improving abilities in ai technology, of human
curiosity, and of industry needs is responsible for the global expansion in
the use of ai technology. This trend continues into the third millennium,
as more traditionally human functions are being replaced by ai technolo-
gies. For example, the South Korean government now uses ai robots as
soldiers to guard the border with North Korea; as teachers in schools; and
since 2012, as prison guards.16
In an eighty-two-page report that outlines the future use of drone air-
craft, “Unmanned Aircraft Systems Flight Plan 2009 – 2047,” the US Air
Force writes that autonomous drone aircraft are key “to increasing effects
while potentially reducing cost, forward footing, and risk.” The report
states that, similar to a chess program that can outperform proficient chess
masters today, future drones will be able to react faster than human pilots
ever could. But the report also notes potential legal problems: “Increasingly
humans will no longer be ‘in the loop’ but rather ‘on the loop’ — monitoring
the execution of certain decisions . . . Authorizing a machine to make lethal
combat decisions is contingent upon political and military leaders resolv-
ing legal and ethical questions.”17
1.1.2. Defining A rtificial I ntelligence, and the E ndlessness

of the Quest for M achina Sapiens
Since the beginning of ai , researchers have been trying to develop com-
puters that actually “think.” This is the holy grail of ai research.18 But to
develop a thinking machine, it is necessary to first define what exactly
thinking is. Defining thinking, with regard to both humans and machines,
proved to be a complicated task for ai researchers. The development of
machines that have the ability to think independently would be an impor-

tant event for humankind, which has claimed a monopoly over this high
mental skill. The creation of a true thinking machine would be tantamount
to the emergence of a new species on earth, the machina sapiens.
But does human science want to create a new species? Since time im-
memorial, people have been trying to play God, with the first preventing
steps mentioned in the Bible.19 Some aspects of modern scientific research
contain elements of similar conceit, including such endeavors as human
cloning, biomedical engineering, anti-aging pursuits, and more. The quest
for the creation of a new species matches this trend. Although creating a
new species may benefit humans, this is not necessarily the motivation
behind ai research. The reason may be much deeper, touching on the most
profound of human quests, which, according to the Bible, was prohibited
to humans following the original sin.
The first move toward the creation of machina sapiens is to define artifi-
cial intelligence. Various definitions have been proffered. Bellman defined
it as “the automation of activities that we associate with human thinking,
activities such as decision-making, problem solving, learning.”20 Hauge-
land defined it as “the exciting new effort to make computers think . . .
machines with mind, in the full and literal sense.”21 Charniak and McDer-
mott defined it as “the study of mental faculties through the use of com-
putational models.”22 Schalkoff defined it as “a field of study that seeks to
explain and emulate intelligent behavior in terms of computational pro-
cesses.”23 Kurzweil defined it as “the art of creating machines that perform
functions that require intelligence when performed by people.”24 Winston
defined it as “the study of the computations that make it possible to per-
ceive, reason, and act.”25 Luger and Stubblefield defined it as “the branch
of computer science that is concerned with the automation of intelligent
behavior.”26 Rich and Knight defined it as “the study of how to make com-
puters do things at which, at the moment, people are better.”27
At first sight, these definitions seem more confusing than clarifying.
But according to these and other definitions, ai systems may be grouped
into four main categories: those that (1) act like humans, (2) think like
humans, (3) think rationally, and (4) act rationally.28
ai systems that act like humans are validated by the Turing test, pro-
posed by the British mathematician in 1950 to provide an operational defi-
nition of intelligence.29 Turing defined intelligent behavior as the ability
to achieve human-level performance in all cognitive tasks that would pre-
vent a human interrogator from distinguishing between human and ma-
chine behavior. The Turing test involves a human “listening” to a conver-
sation between a machine and a human. The conversation is carried out in

writing to avoid identifying the machine through its voice; therefore, the
“listening” is carried out in reading the text of the questions and answers.
The machine is considered to have passed the Turing test if the human
interrogator cannot tell the difference between the two.30 The Turing test
assumes equal cognitive abilities for all humans, although conversations
between a machine and a child, a person with an intellectual disability, a
very tired person, and the designer of the machine are likely to be signifi-
cantly different.31
To identify ai systems that think like humans, it would be necessary to
define human thinking first. ai technologies developed as general problem
solvers were designed to make decisions that are very similar to human
decisions based on the same information.32 The modern development of
cognitive science enabled experimental approaches to mechanical think-
ing, and to the testing of machine thinking. In another test proposed by
Turing, the interrogator’s objective is to differentiate between a man and a
woman in a conversation, a test that depends on the communicative skills
of the human participants no less than on that of the machine.33
The effectiveness of the original Turing test has been questioned, espe-
cially regarding strong ai (in which computers have a mind similar to that
of humans). Searle, in his famous Chinese Room thought experiment,34
described a person who does not know Chinese, alone in a room, receiv-
ing batches of Chinese messages in writing. A rule book, written in the
person’s native language, allows him or her to look up the Chinese char-
acters by their shape, and then to answer the messages using this rule
book. People outside the room, who are conducting the Turing test, are
convinced that the person inside the room understands Chinese.
In this thought experiment, the person inside the room simply fol-
lows the instructions in the book, but neither the person nor the book un-
derstands Chinese, although both can simulate such understanding. The
instruction book is, in this case, a computer program. What, then, does it
mean for humans to understand a foreign language? It is difficult to iden-
tify an ai system that thinks rationally without defining rationality first. If
rationality means “right” thinking, it can be represented by formal logic,
so that given correct information, the machine can aim to produce a com-
bination of right conclusions. For example, knowing that all tomatoes are
red and that X is a tomato, the machine should reach the conclusion that
X is red. Most modern ai systems support formal logic and know how to
behave in accordance with it.
ai systems that act rationally are an advanced variation of systems that
think rationally. The latter can reach the right conclusions from correct in-
formation as outsiders, but the former can participate in events, taking the
right action based on correct information. For example, an ai goalkeeper
in a soccer game that sees the ball approaching the goal rapidly can not
only calculate the speed and angle at which the ball is approaching and
determine the action required to block it, but also take the required action
and stop the ball.
The quest for machina sapiens, however, reaches much deeper than
these classification and definitions, and if the quest is successful, it will
answer the question of what makes humans intelligent, and will result in
the design of intelligent machines. Since the late 1980s, the accepted ap-
proach has been that intelligent thinking is identified by certain attributes,
specifically the following five: (1) communication, (2) internal knowledge,
(3) external knowledge, (4) goal-driven conduct, and (5) creativity.35 Let us
discuss each of these attributes.
Communication is considered the most important attribute that defines
an intelligent entity, because we can communicate with intelligent enti-
ties. Note that we communicate not only with humans, but with certain
animals as well, although the range of such communication is narrower
than our range with humans, and not all ideas can be expressed. For exam-
ple, you can let your dog know how angry you are, but you cannot explain
to him the Chinese Room thought experiment. This situation is not differ-
ent from communicating with a two-year-old human. The more intelligent
the other entity, the more complex the communication conducted with it.
Communication assumes proper understanding of the information
contained in it, and can be used to test the ability to understand compli-
cated ideas. But communication does not always attest to the quality of
understanding. Some highly intelligent people, at genius level, are not
good communicators, and some autistic geniuses cannot communicate at
all. At the same time, most normative people have advanced communica-
tion skills, although only a few can communicate about complex matters
such as Einstein’s general theory of relativity, quantum mechanics, parallel
universes, or exponential equations.
Communication can be carried out not only through speech, but by
other means as well, including writing. Therefore, machines can be quite
intelligent even if they lack the ability to speak, similar to mute people. So
although the communication attribute of intelligence is important, there
are many exceptions to it. The Turing test is based on this attribute, but if
human communication is so inaccurate, how can we trust it to identify ai ?
Internal knowledge refers to the knowledge of entities about them-
selves. This parallels self-awareness. An intelligent entity is supposed to

know about its own existence, about the fact that it functions in some way,
about the way in which it integrates into factual reality, and so on. Formal
logic reasoning showed the way to artificial internal knowledge through
self-reference.36 It is possible to program computers to appear that they
know about themselves, and that they know about knowing about them-
selves. But many researchers consider this too artificial, and insist that
it would be difficult to determine whether these computers really know
about themselves. Yet no alternatives have been suggested to test internal
knowledge. The question is, how is it possible to know anything about
another person’s internal knowledge?
External knowledge refers to factual data about the outside world and
about factual reality. This attribute is considered crucial in an age when
knowledge functions as a commodity, especially with relation to expert
systems.37 An intelligent entity is expected to know how to find data about
the outside world, and to know the facts that make up the factual reality to
which it is exposed. This attribute assumes memory skills and the ability
to classify information into seemingly relevant categories. This is the way
humans assemble their life experience and the way in which they learn. It
is difficult to act as an intelligent entity if all factual elements are treated
each time as new. Although factual events are new each time, they do
contain common characteristics that an intelligent entity should identify.
For example, medical expert systems, which are designed to diagnose
diseases based on their symptoms, should be able to identify the common
characteristics of diseases in many cases, even if these cases vary to an
extreme degree. An entity that lacks this ability acts in a way similar to
people who suffer from amnesia. Although they conduct themselves ade
quately in a given circumstance, their conduct is not added to their cu-
mulative experience. This is how simple machines operate: performing a
certain task, but not knowing that they have done so, and not having the
ability to draw on this or other experiences to guide them in future tasks.
So the question is, if inexperienced people can still be considered intel-
ligent, then why not machines that behave similarly?
Goal-driven conduct marks the difference between random or arbitrary
conduct and intended conduct. Goal-driven conduct requires an operative
plan to achieve the desired goal. For most humans, goal-driven conduct
is interpreted as intention. If one is hungry and sees an apple, eating it is
goal-driven conduct, the goal being to appease the hunger. We may say that
the person ate the apple with the intention of not being hungry anymore.
Goal-driven conduct is not unique to humans. When a dog sees a juicy
bone behind an obstacle, she plans to bypass the obstacle to get the bone.
Executing the plan is goal-driven conduct on the part of the dog.
Different creatures may have different goals, of differing levels of com-
plexity. The more intelligent the entity, the more complex are its goals.
The goals of dogs are at the level of calling for help when their masters are
in distress, whereas the goals of humans may include landing on Mars.
Computers have the ability to plan the landing on Mars, and robots and
unmanned vehicles are already exploring Mars, searching for data. The
reductionist approach to goal-driven conduct deconstructs a complex goal
into many simple ones; achieving these is considered goal-driven conduct.
Goals and plans for achieving them may be incorporated into computer
programs. But not all humans pursue complex goals at all times and under
all circumstances. The question is, what level of complexity of goals is
needed for a machine to be considered intelligent?
Creativity involves finding new ways of understanding or doing things.
All intelligent entities are assumed to have some degree of creativity. When
a fly tries to get out of the room through a closed window, it crashes against
the windowpane over and over again, trying invariably the same conduct
without success. This is the opposite of creativity. At some point, however,
the fly does seek another way out, which is considered to be more creative.
A dog may find a way out in much fewer attempts; a human in fewer still.
Computers may be programmed not to repeat the same conduct more than
once, and instead to seek other ways of solving the problem. This type of
behavior is essential for general problem-solving software.
Creativity is not homogenous, and has degrees and levels. Not all humans
are considered to be thinking outside the box, and many perform their daily
tasks exactly the same way day after day. Most lottery players play the same
numbers time after time for years without winning. What makes their crea
tivity different from that of the fly? The fly’s chance of breaking through the
window is not smaller than that of the lottery player winning the main prize.
Many factory workers perform the same operations for hours every day, and
they are considered intelligent entities. The question is, what is the exact
level of creativity required to be considered intelligent?
Not all humans who are considered intelligent share all five attributes.
Why, then, should we use different standards for humans and machines to
measure intelligence? But each time some new software has succeeded in
conquering a given attribute, the achievement was rejected for not being
real communication, internal knowledge, external knowledge, goal-driven
conduct, or creativity.38 Consequently, new tests were proposed to make
sure that the relevant ai technology was really intelligent.39
Some of these tests have had to do with knowledge representation

(what the machine knows), decoding (translating knowledge from factual
reality to its representation), inference (extracting the content of knowl-
edge), control of combinatorial explosion (preventing endless calculation
for the same problem), indexing (arranging and classifying knowledge),
prediction (assessing the probability of possible factual events), dynamic
modification (self-changing the programs as a result of experience), gener-
alization (inductive interpretation of factual events), and curiosity (won-
dering about factual events or seeking reasons for them).
In their biological sense, all the attributes just mentioned are manifes-
tations of the human brain. Nobody doubts this fact. These attributes are
manifestations of neuronal activity in the brain, which can be identified
and quantified. If so, and if transistors can be functionally activated in the
same way as neurons, why can these attributes not be reproduced by means
of transistors?40 The simple answer is that ai entities are simply not human.
The process of proposing new tests whenever a given ai technology
succeeds in passing tests devised earlier makes the quest for machina sa-
piens endless. The reason for this is psychological rather than technologi-
cal. When talking about ai , we must not envisage regular human beings
with a robotic-metallic appearance. Yet we will settle for nothing less,41
forgetting that artificial intelligence happens to be artificial, not natural.
When ai technology finally succeeds in passing a certain test, this shows
us that the problem was not in the technology, but in the test. Because the
complexity of the human mind is too vast to be tested with simple tests,
we replace the tests, a process in which we have learned more about the
human mind, especially about the “bureaucracy” of mind and its inten-
tionality,42 than about technology.43
The template of the arguments against the definition and feasibility of
machina sapiens as a machine that possesses truly intelligent personhood
goes like this:
1. To possess personhood, an entity must have attribute X.

2. a i technology cannot possess attribute X.
3. Behavior of a i technology that is identified to possess attribute
X demonstrates that it can simulate or imitate that attribute.
4. Simulation of attribute X is not attribute X itself.
Therefore, ai technology is not really intelligent. Some scholars called

this template of arguments a “hollow shell strategy.”44 Note that argument
number 2 and the conclusion result in a catch-22: ai entities are not intel-
ligent because ai technology cannot possess the attribute of intelligence.
The contents of attribute X may be made up entirely of advanced in-

telligence tests, which in this context reflects the belief that intelligence
can be only human, characterized by the fact that it is exclusively human,
so that machines cannot perform it. The result is a paradox. Despite re-
markable advances in ai technology, there has been great frustration with
the abilities of ai , as progress in ai research has underscored how far the
technology is from imitating the human mind and how much farther the
mysteries of the human mind still need to be explored.
In the 1950s, it was difficult to believe that a computer could ever
defeat humans in chess, and that if it were to happen, computers would
be considered truly “intelligent.” But in 1997 a computer program (Deep
Blue) defeated the world champion in chess, and it was still not consid-
ered intelligent.45 In 2011 another computer program (Watson) defeated
two of the top Jeopardy players on television, but it was still not consid-
ered intelligent.46 Although the computer missed some of the jokes on the
show, it won the contest. The technology developed for Watson is being
adapted for use in advanced expert systems, and although it is considered
good enough to diagnose diseases, many people still do not consider it
intelligent.
ai technologies pioneered machine learning, an inductive method of
learning in which the computer analyzes various cases and situations sub-
mitted to it, and by means of generalizations, produces an image of the
facts for future use.47 Similar knowledge in humans is called expertise,
but humans prefer not to refer to these computer systems as intelligent.
It seems that the creation of machina sapiens recedes further with each
advance in ai technology, and that ai research has produced an endless
quest for the new species.
There are two ways of dealing constructively and positively with the
endless quest for machina sapiens. The first is through technological re-
search, seeking ways to reduce the gap between machines and humans.
This is the path chosen by ai researchers since 1950, following the orig-
inal quest for machina sapiens in the belief that some day technology
will be able to imitate the human mind. ai research has had significant
accomplishments by following this method.48 The second way is indus-
trial. As I will describe, industry is not interested in imitating the human
mind in all its details. Indeed, industry welcomes the opportunity of using
ai entities that do not exhibit certain human qualities. So the disadvan-
tages of machines — the limited ways in which they imitate the human
mind — become advantages for industry. Turning the disadvantages of the
machine into advantages for industry makes it possible to increase the use
and integration of ai entities in industry. The industrial use of ai technol-

ogy was the catalyst for the emergence of machina sapiens criminalis.
1.1.3. Turning Disadvantages into A dvantages:

I ndustrial and P rivate Use of A rtificial I ntelligence
Industrial use of artificial intelligence technology is not new. As noted ear-
lier,49 ai technology has been embraced in advanced industries since the
1970s. But whereas in the beginning, ai technology was used by industry
because of its similarity to the human mind, later it was used because of its
differences from the human mind. Industry was quick to understand that
complete and perfect imitation of the human mind would not be as useful
as incomplete imitation, so industry encouraged the development of ai
technology as long as imitation of the human mind was not complete. And
because complete imitation of the human mind is still far in the future,
industry and ai research continue to cooperate.
How did industry turn a disadvantage into an advantage? Consider the
calculator. If we enter “1 3 1 =” into a calculator repeatedly, we continue
to receive the answer “1” invariably, even if we repeat this operation thou-
sands of times. The process activated by the calculator is the same each
time. But if we ask a human the same question, we may receive an answer
the first time — unless the human believes she is being teased — and per-
haps several more times, but not thousands of times. At some point, the
human stops answering because she becomes bored, irritated, or nervous,
or because she loses any desire to answer.
This is a major problem from the point of view of ai researchers, and it
underscores the fact that the human mind may act arbitrarily, sometimes
for irrational reasons. It would be unthinkable, however, for the calcula-
tor to refuse to answer “1 3 1 =” even for the thousandth time. For tasks
of this type, industry prefers entities that are not bored with our requests
or whims, are not irritated by our questions, and will continue to serve us
well, no matter how many times we ask the same question.
It appears that most humans lack these abilities because they have a
human mind. But machines, which have not succeeded in imitating the
human mind completely, have the ability to provide us with this kind of
service. The example just presented is theoretical, of course, as no one
would really type a thousand times “1 3 1 =” on one’s calculator, a task
that in itself would require nonhuman skills. But machine skills of this
nature are required in most industries.
Consider the customer service representatives of large companies that
serve hundred of thousands of customers and must provide polite, help-
ful answers to requests of every type. How do these representatives re-

spond after one call? After a hundred calls? After thousands of calls? How
does the number of calls affect the quality of service? But the technologi-
cal inability of a machine to become bored, irritated, or tired is a net ad-
vantage for industry. An automatic customer-service system services the
thousandth customer exactly the way it serviced the first one: politely,
patiently, efficiently, and accurately.
Expert systems specializing in medical diagnosis are not bored by re-
peating identical solutions to problems of different patients. Police robots
are not too frightened to dismantle dangerous explosive devices. Factory
robots are not bored by performing identical activities thousands of times
a day. Industry has leveraged the nonhuman qualities of ai entities, turn-
ing the traditional disadvantages of ai technology (at least from the point
of view of ai researchers) into advantages for modern industry.50
In practice, these disadvantages are benefitting not only industry. Re-
search in ai technology and industry has made this technology available
for private consumption as well. Personal robot assistants based on ai tech-
nology are already available, and ai robots are expected to make their ap-
pearance in family and private life, including some of the most intimate
situations. It has already been suggested that “love and sex with robots”
may provide a better and socially healthier alternative to prostitution.51
No shame, abuse, or mental or physical harm would result from using the
ai alternative, and this could result in a real social change.
Household robots are not offended if they are asked to perform the
same action over and over, and they do not require vacations, demand
raises, or ask for favors. Teacher robots are unlikely to teach material out-
side the curriculum they have been programmed to teach, and prison-
guard robots are unlikely to be bribed to allow prisoners to escape. These
nonhuman qualities have made ai technology highly popular for meeting
industrial and private needs.
We can characterize the ai technology required for industrial and pri-
vate needs as a less than complete and perfect imitation of the human
mind, but as having some human skills and being capable of an imperfect
and incomplete imitation of the human mind. The resulting ai entities are
not yet machina sapiens, but they possess some human skills for solving
problems and can imitate some of the capabilities of the human mind. The
existing capabilities of ai technology, which are already being used to sat-
isfy industrial and private needs, are incorporated today in the machina
sapiens criminalis.
1.2. On Evolu tion and Devolu tion :

Machin a Sapiens C r iminalis
In modern times, technology and law often find themselves on opposite
sides of the spectrum. Whereas technology serves the cause of the evolu-
tion of new innovations, the law pulls in the conservative direction. For
example, technology made human cloning possible in principle, yet in
most countries, the law prohibited it or initiated a public debate about
it. What technology considers progress, the law may treat as regress. At
times the law is inspired by moral conceptions of society, at other times
by human fears. This is the situation with advanced robotics based on ai
technology that has produced the machina sapiens criminalis.
1.2.1. Serve and P rotect: H uman F ears of H uman-R obot C oexistence

Since the beginning of the twenty-first century, reports about research
in advanced technology have predicted that starting at about the third
or fourth decade of this century, a new generation of robots will support
human-robot coexistence.52 Under the Fukuma “World Robot Declaration”
issued in 2004, these robots are anticipated to coexist with humans, assist
humans both physically and psychologically, and contribute to the realiza-
tion of a safe and peaceful society.53
It has been accepted that there are two major types of such robots.
The first is a new generation of industrial robots capable of manufacturing
a wide range of products, performing multiple tasks, and working with
human employees. The second is a new generation of service robots capa-
ble of performing such tasks as house cleaning, security, nursing, life sup-
port, and entertainment by coexisting with humans in home and business
environments. The authors of most published projections on this technol-
ogy have also evaluated the level of danger to humans and society as a
result of using these robots.54 Such evaluations provoked a debate about
the safety of using robots, irrespective of the real level of predicted danger.
People consider safety only when something is thought dangerous.
The accelerated technological developments in the area of ai and robotics
caused many fears. For example, one of the first reactions to a humanoid
robot as a caregiver designed to provide nursing care was the fear of hurt-
ing the assisted human. Would all humans be ready to let their babies and
children be nursed by robots? Most humans are not experts in technology,
and most fear what they do not know.
Further, the vacuum created by the absence of knowledge and certainty
is sometimes filled by science fiction. In the past, science fiction was rare
and had a small following. Today, Hollywood brings science fiction to

people worldwide. Most blockbusters of the 1980s, 1990s, and twenty-first
century are classified as science fiction movies, and most of these films ex-
ploit people’s fears. For example, in 2001: A Space Odyssey (1968), based
on Arthur C. Clarke’s novel, the central computer of the spaceship is out
of human control and attempts to assassinate the crew.55 Safety is restored,
after many deaths, only when the computer is shut down.56
In The Terminator series of films (The Terminator, 1984; Terminator
2: Judgment Day, 1991; and Terminator 3: Rise of the Machines, 2003),
machines are taking over humanity, which is almost extinct. A few sur-
vivors establish a resistance to oppose the machines, which must be shut
down to ensure human survival. Even the savior, which happens to be a
machine, ends up shutting itself down.57 In the trilogy of The Matrix (The
Matrix, 1999; Matrix Reloaded, 2003; and Matrix Revolutions, 2003), ma-
chines dominate the earth, enslaving humans, to produce energy for their
own benefit. The machines control the humans’ minds by creating the il-
lusion of a fictional reality called the “matrix.” Only a few can escape the
matrix to fight for freedom from the domination of the machines.58 In I,
Robot (2004), based on an Isaac Asimov collection of stories from 1950, an
advanced model of robots hurts people, one robot is suspected of murder,
and one of the human heroes is a detective who does not trust robots. The
overall plot concerns the robots’ attempt to take over humans.59
The influence of Hollywood is vast, so it is reasonable to expect that
fear of artificial intelligence and robotics will dominate the public mind.
The more advanced a robot, the more dangerous. A popular topic in sci-
ence fiction literature and films is revolt and takeover by robots. Thus,
when people think about advanced robots and technology, in addition to
utility and unquestionable advantages, they also consider how to protect
themselves. People will accept the idea of human-robot coexistence only
if they feel safe from the other, the different, the terrifying: the machine.
What protection mechanisms can humanity use to ensure the safety of
coexistence with robots and ai technology?60 The first to sound the alarm
about the dangers of such technology was science fiction literature, which
was also the first to suggest protection from it. The first circle of protec-
tion to be proposed was in the realm of ethics focused on safety. The ethi-
cal issues were presented to the designers and programmers of ai entities,
to ensure that built-in software would prevent any unsafe activity of the
technology.61 One of the pioneers of the attempts to create robot ethics was
Isaac Asimov.
Asimov stated three famous “laws” of robotics in his collection of nine
science fiction stories assembled under the title I, Robot in 1950: “(1) A
robot may not injure a human being or, through inaction, allow a human
being to come to harm; (2) A robot must obey the orders given it by human
beings, except where such orders would conflict with the First Law; (3) A
robot must protect its own existence, as long as such protection does not
conflict with the First or Second Laws.” 62 In the 1950s, these laws were
considered innovative, and restored some confidence in a terrified public.
After all, harming humans was not allowed.
Asimov’s first two laws represent a human-centered approach to safety
in relation to ai robots. They also reflect the general approach that as robots
gradually take on more intensive and repetitive jobs outside industrial
factories, it is increasingly important that safety rules support the concept
of human superiority over robots.63 The third law straddles the border be-
tween human-centered and machine-centered approaches to safety. The
functional purpose of the robots is to satisfy human needs, and therefore
robots should protect themselves in order to better perform these tasks,
functioning as human property.
But these ethical rules, or laws, were insufficient, ambiguous, and not
sufficiently broad, as Asimov himself admitted.64 For example, a robot in
police service trying to protect hostages taken by a criminal who intends
to shoot one of them understands that the only way to stop the murder of
an innocent person is to shoot the criminal. But the first law prohibits the
robot from killing or injuring the criminal (by acting), and at the same time
prohibits it from letting the criminal kill the hostage (through inaction).
What can we expect the robot to do based on the first law, with no other
alternatives available? Indeed, what would we expect a human police of-
ficer to do? Any solution breaches the first law.
Examining the other two laws does not change the consequences. If
a human commander ordered the robot to shoot the criminal, the robot
would still be breaching the first law. Even if the commander were in im-
mediate danger, it would be impossible for the robot to act. And even if
the criminal threatened to blow up fifty hostages, the robot would not be
allowed to protect their lives by injuring the criminal. Matters would be
even more complicated if the commander were a robot too, but the con-
sequences would be no different. And the third law would not alter the
results even if the robot itself were in danger.
The dilemma of the police robot is not uncommon in a society in
which humans and robots coexist. Any activity by robots and ai technol-
ogy in such a society is riddled with such dilemmas. Consider a robot in
medical service that is required to perform an intrusive emergency surgi-
cal procedure intended to save a patient’s life, and to which the patient
objects. Any action or inaction by the robot breaches Asimov’s first law,
and any order by a superior is unable to solve the dilemma because an
order to act causes injury to the patient and an order to refrain from action
results in the patient’s death.
Some dilemmas are easier to solve when one of the options involves no
injury to humans, but this raises other questions. In the previous example
of robots used as prison guards,65 what should such a robot do when a pris-
oner attempts to escape and the only way to stop the prisoner is by causing
injury? What should a sex robot do when ordered to commit sadistic acts?
If the answers are not to act, then the question is, why use such robots in
first place if their intended actions breach the first law?
If humanity determines that the destiny of ai robots and other artifi-
cial entities is to serve and protect humanity in various situations, this
involves difficult decisions that will have to be made by these entities.
The terms injury and harm may be wider than specific bodily harm, and
these entities may harm people in other ways than bodily. Moreover, in
various situations, causing one sort of harm should be necessary in order
to prevent greater harm. In most cases, such decision involves complicated
judgment that exceeds the simple dogmatic rules of ethics.66
The debate about Asimov’s laws gave rise to debate about the moral ac-
countability of machines.67 The moral accountability of robots, machines,
and ai entities is part of the endless quest for machina sapiens. But robots
exist, they participate in daily human life in industrial and private envi-
ronments, and they do cause harm from time to time, regardless of whether
they can be subjected to moral accountability. Therefore, the ethical sphere
is unsuitable for settling the issue. Ethics involve moral accountability,
complicated inner judgment, and inapplicable rules. The question, then,
is what can settle robot functioning, human-robot relations, and human-
robot coexistence? The answer may lie in criminal law.
1.2.2. E volution and Devolution:

M achina Sapiens C riminalis as a By-product
Artificial intelligence technology makes possible a range of industrial and
private uses of robots. The technology will certainly become more ad-
vanced in the future as ai research develops over time, and both indus-
trial and private use of this technology will broaden the range of tasks that
ai robots will undertake. The more advanced and complicated the tasks,
the higher the chances of failure in accomplishing them. Failure is a broad
term that includes various situations in this context. The most common
situation is that the task undertaken by the robot has not been accom-
plished successfully. But some failure situations can involve harm and
danger to individuals and society.
For example, the task of prison-guard robots has been defined as pre-
venting escape by using minimal force against the prisoners. A prisoner
attempting to escape may be restrained by the robot guard, which holds
the prisoner firmly but causes injury; the prisoner may then argue that the
robot has excessively used its power. Analyzing the robot’s actions may
reveal that it could have chosen a more moderate action, but the robot had
evaluated the risk as being graver than it actually was. In this case, who is
responsible for the injury?
This type of example raises important questions and many arguments
about the responsibility of the ai entity. If analyzed through the lens of
ethics, the failure in this situation is that of the programmer or the de-
signer, as most scientists would argue, not of the robot itself. The robot
cannot consolidate the necessary moral accountability to be responsible
for any harm caused by its actions. According to this point of view, only
humans can consolidate such moral accountability. The robot is nothing
but a tool in the hands of its programmer, regardless of the quality of its
software or cognitive abilities. This argument is related to the debate about
machina sapiens.
Moral accountability is indeed a highly complex issue, not only for
machines, but for humans as well. Morality, in general, has no common
definition that is acceptable in all societies by all individuals. Deonto-
logical morality (concentrated on the will and conduct) and teleological
morality (concentrated on the result) are the most acceptable types, and in
many situations they recommend opposite actions.68 The Nazis considered
themselves deontologically moral, although most societies and individu-
als disagreed. If morality is so difficult to assess, then moral accountability
may not be the most appropriate and efficient way of evaluating responsi-
bility in the type of case we have just examined.
In this context, the issue of the responsibility of ai entities will always
return to the debate about the conceptual ability of machines to become
human-like, so that the endless quest for machina sapiens would become
an endless quest for ai accountability. The relevant question here exceeds
the technological one, and it is mostly a social question. How do we, as a
human society, choose to evaluate responsibility in situations of harm and
danger to individuals and society?
The main social tool available for handling such situations in daily
life is criminal law, which defines the criminal liability of individuals
who harm society or endanger it. Criminal law also has educational social
value because it educates individuals on how to behave within their so-
ciety. For example, criminal law prohibits murder; in other words, the
law defines what is considered to be murder, and prohibits it. This has
the value of punishing individuals for murder ex post, and prospectively
educating individuals not to murder ex ante, as part of the rules of living
together in society. Thus, criminal law plays a dominant role in social
control.69
Criminal law is considered the most efficient social measure for di-
recting individual behavior in any society. It is far from perfect, but under
modern circumstances it is the most efficient measure. And because it is
efficient regarding human individuals, then it is reasonable to examine
whether it is efficient regarding nonhuman entities as well, specifically
ai entities. Naturally, the first step in evaluating the efficiency of criminal
law with regard to machines is to examine the applicability of criminal
law to them. So the question in this context is whether machines may be
subject to criminal law.
Because criminal law does not depend on moral accountability, the
debate about the moral accountability of machines is irrelevant to the
question at hand. Although at times there is an overlap between criminal
liability, certain types of moral accountability, and certain types of ethics,
such coincidence is not necessary in order to impose criminal liability.
The question of the applicability of criminal law to machines involves two
interrelated, secondary questions:
1. Is criminal liability applicable to machines?

2. Is criminal punishment applicable to machines?
The first question is a core issue of criminal law, and imposition of crimi-
nal liability depends on meeting its requirements. Only if these require-
ments are met does the question of punishability arise, that is, how human
society can punish machines. The answers proposed in this book are affir-
mative for both questions. Chapters 2 through 5 examine the first question,
and chapter 6 examines the second. If the affirmative answers offered in
this book are accepted, then a new social entity has indeed been created,
machina sapiens criminalis.
Machina sapiens criminalis is the inevitable by-product of human ef-
forts to create machina sapiens. Technological response to the endlessness
of the quest for machina sapiens has resulted in advanced developments
in ai technology, which have made it possible to imitate human minds
and their associated skills better than ever before. Current ai technology
is capable of performance that was considered science fiction only a few

years ago — a situation that has existed over the entire second half of the
twentieth century, and continues into the twenty-first century.
Every step along the road of technological development is another step
in the evolution of machina sapiens. The endless quest for machina sa-
piens created very high upper and lower thresholds, so high in fact that
today, it is necessary to be human to be considered machina sapiens. So
the quest remains endless, and the technological race continues. But crim-
inal liability does not necessarily require the presence of all human skills.
To become an offender, one need not use all human skills, regardless of
whether one possesses such skills.
Consider the attribute of creativity, discussed earlier.70 Naturally, there
are many types of thieves. Some are deemed creative, others not. When
examining their criminal liability for a given offense, their human skills of
creativity are not even considered. We impose criminal liability for theft
regardless of whether the offender was creative in committing the theft.
Because creativity is not a condition for the imposition of criminal liability
on any offender, it is not considered in the legal process.
When the requirements of criminal liability for the specific offense are
met, no other qualifications, skills, or thoughts are considered. It is pos-
sible to argue that although theft does not require creativity, humans are
assumed to be creative, and therefore criminal liability is applicable to
them. This type of argument may be relevant in the debate on the endless-
ness of the quest for machina sapiens, but absolutely not in the question
of criminal liability. Human skills that are irrelevant for the commission
of a given offense are not considered as part of the legal process toward
criminal liability.
Therefore, whether the offender has been creative in committing the
offense is irrelevant for the imposition of criminal liability, because there
is no such requirement for criminal liability. Whether the offender was or
was not creative outside the commission of the offense is also irrelevant,
because the legal process focuses only on the facts involved in the com-
mission of the offense. This is also true of many other attributes consid-
ered necessary for announcing the birth of machina sapiens.
It appears, therefore, that in the race to create machina sapiens, signifi-
cant by-product has been developed: a type of machine that does not have
the skills to be considered machina sapiens because it cannot fully imitate
the human mind; but its skills are adequate for various activities in indus-
trial and private environments. Indeed, the absence of some skills is even
considered an advantage when focusing on industrial and private use,
although ai research considers such absence a disadvantage.71 One type of

activity that this by-product is capable of is the commission of offenses.
In other words, although the by-product is not capable of many types
of creative activities, it is perfectly capable of committing offenses. The
reason for this lies in the definitions and requirements of criminal law for
the imposition of criminal liability and punishment. These requirements
are satisfied by capabilities that are far lower than those required to create
machina sapiens. In the context of criminal law, as long as these require-
ments are met, there is nothing to prevent the imposition of criminal liabil-
ity, whether the subject of criminal law is human or not. This is the logic
that applies to human and nonhuman offenders, such as corporations.
Thus, as long as all relevant requirements of criminal law are met, a
new type of subject may be added to the large group of existing subjects to
criminal law, in addition to human individuals and corporations. These
subjects may be referred to as machina sapiens criminalis, not a subtype of
a general machina sapiens, but its by-product. This by-product is consid-
ered to be less advanced technologically, because it is possible to belong
to this category of machines without the high-level skills attributed to the
ideal machina sapiens, if and when achieved.
From the technological-scientific point of view, any progress toward
the development of machina sapiens is considered progress (or “evolu-
tion”),72 whereas the by-product, machina sapiens criminalis, is more
of a regress (or “devolution”). Machina sapiens criminalis is a stopping
point along the race to the top. It is a less-advanced peak, because the re-
quirements for criminal liability are much lower than those for the ideal
machina sapiens. Nevertheless, machina sapiens criminalis is not a race
to the bottom, but merely a stopping point in the race to the top. From
the criminal law point of view, the technological research may rest at this
point because the entities to which the criminal law is applicable already
exist.
If the range of machines extends between the poles of technically basic
(“dumb”) machines and machina sapiens, then machina sapiens crimi-
nalis is situated somewhere in the middle, perhaps closer to the machina
sapiens pole. But as noted earlier, the emergence of machina sapiens
criminalis is contingent on a match between the requirements of criminal
law and the relevant skills and abilities of the machine. Some researchers
argue that the current law is inadequate for dealing with ai technology,
and that it is necessary to develop a new legal domain called Robot Law.73
The argument of this book is that the current criminal law is adequate
to cope with ai technology. Moreover, if technology were to advance sig-
nificantly toward the creation of machina sapiens, this would make cur-
rent criminal law even more relevant to addressing ai technology because
such technology imitates the human mind, and the human mind is already
subject to current criminal law. Thus, the closer ai technology approaches
to a complete imitation of the human mind, the more relevant the current
criminal law becomes.
Subjecting ai robots to the criminal law may relax our fears of human-
robot coexistence. Criminal law plays an important role in giving people
a sense of personal confidence. Each individual knows that all other indi-
viduals in society are bound to obey the law, especially the criminal law.
If the law is breached by any individual, society enforces it by means of
its relevant coercive powers (police, courts, and so on). If any individual
or group is not subject to the criminal law, the personal confidence of the
other individuals is severely harmed because those who are not subject to
the criminal law have no incentive to obey the law.
The same logic works for humans, corporations, and ai entities.74 If any
of these is not subject to criminal law, the personal confidence of the other
entities is harmed, and more broadly, the sense of confidence of the entire
society is harmed. Consequently, society must make all efforts to subject
all active entities to its criminal law. At times this requires conceptual
changes in the general insights of society. For example, when the fear of
corporations that are not subject to criminal law became real, in the seven
teenth century, criminal law was applied to them. It is perhaps time to do
the same for ai entities in order to alleviate human fears.75
But one more issue must be clarified. There are other creatures living
among us, in our society, in coexistence with us: animals. Why not apply
our legal rules concerning animals to ai entities as well?
1.2.3. I napplicability of the Zoological

L egal M odel to ai Technology
Not all creatures on earth are considered by humans to be intelligent. Most
of those considered to be intelligent are mammals. Nevertheless, humans
and animals have coexisted for many millennia. Since humanity first do-
mesticated animals for its own benefit, this coexistence has become inten-
sive. We eat their products (milk, eggs, and so on); feed on their flesh; are
protected by them (for example, guard dogs); employ them (in agriculture,
transportation, and so on); and use them for both industrial and private
needs. Therefore, the law has addressed the issues of our coexistence since
ancient times.76
Examination of both ancient and modern law regarding animals re-
veals two important aspects that the law addresses. The first is the law’s
relation to animals as a property of humans, and the second is our duty
to show mercy toward animals. The first aspect involves the ownership,
possession, and other property rights of humans toward animals. For ex-
ample, if damage is caused by an animal, the human who has the property
rights over the animal is legally responsible for the damage. If a person is
attacked by a dog, its owner is legally responsible for any damage.
In most countries these are issues of tort law, although in some coun-
tries they are related to criminal law as well, but in either case the legal
responsibility is the human’s, not the animal’s. Since ancient times, an
animal was incapacitated only if it was considered too dangerous for so-
ciety. Incapacitation usually meant killing the animal. This was the case
if an ox gored a person in ancient times,77 and it still is the case if a dog
bites a person under certain circumstances. No legal system considers an
animal to be directly subject to the law, especially not to criminal law,
regardless of the animal’s intelligence.
The second aspect of law — the duty to show mercy toward animals —
is aimed at humans. As humans are considered superior to animals
because of their human intelligence, animals are viewed as helpless. Con-
sequently, the law prohibits the abuse of power by humans against animals
because of cruelty. The subjects of these legal provisions are humans, not
animals. Animals abused by humans have no standing in court, even if
damaged. The legal “victim” in these cases is society, not the animals, so
in most cases, these legal provisions are part of criminal law.
Society indicts humans who abuse animals because such cruelty
harms society, not the animals. This type of protection of animals differs
little from the protection of property. Most criminal codes prohibit dam-
aging another’s property, in order to protect the property rights of the pos-
sessor or owner. But in the case of cruelty to animals, this type of protec-
tion has nothing to do with property rights. The legal owner of a dog may
be indicted for abusing the dog regardless of property rights with respect
to the dog. These legal provisions, which have existed since time imme-
morial, form the zoological legal model. The question is why this model
cannot be applied to ai entities.
We are considering three types of entities: humans, animals, and artifi-
cially intelligent entities. If we wish to subordinate ai entities to the crimi-
nal law of humans, we must be able to justify the resemblance between
humans and ai entities in this context. Indeed, we should explain why
ai entities carry more resemblance to humans than to animals. Otherwise,
the zoological legal model would be satisfactory and adequate for settling
the ai activity. Thus, the question is, who does an ai robot resemble more:
a human or a dog?
The zoological legal model has been examined previously with re-
spect to ai entities when discussing the control of unmanned aircrafts78
and other machines,79 including “new generation robots.”80 For some of
the legal issues raised in these contexts, the zoological legal model was
able to provide answers, but the core problems could not be solved by this
model.81 When the ai entity was able to figure out by itself, using its soft-
ware, how it needed to act, something in the legal responsibility puzzle
was still missing. Communication of complicated ideas is much easier
with ai entities than with animals. This is the case regarding external
knowledge and the quality of reasonable conclusions in various situations.
An ai entity is programmed by humans to conform to formal human
logical reasoning. This forms the core reasoning on which the activity
of the ai entity is based, so that its calculations are accountable through
formal human logical reasoning. Most animals, in most situations, lack
this type of reasoning. It is not that animals are not reasonable, but their
reasonableness is not necessarily based on formal human logic. Emotion-
ality plays an important role in the activity of most living creatures, both
animals and humans, and it may supply the drive and motivation to some
human and animal activity. This is not the case with machines.
If emotionality is used as a measure, humans and animals are much
closer to each other than either is to machines. But if pure rationality is
used as a measure, machines may be closer to humans than to animals.
Although emotionality and rationality affect each other, the law distin-
guishes between them regarding their applicability. For the law, especially
for criminal law, the main factor being considered is rationality, whereas
emotionality is only rarely taken into account. For example, a murderer’s
feelings of hatred, envy, or passion are immaterial for his or her convic-
tion; and only in rare cases, when the provocation defense is used, may
the court consider the emotional state of the offender.82
Further, the zoological legal model educates humans to be merciful
toward animals, as already noted, which is a key consideration under this
model. In relation to ai entities, however, this consideration is immaterial.
Because ai entities lack basic attributes of emotionality, they cannot be sad-
dened, made to suffer, disappointed, or tortured in any emotional manner,
and this aspect of the zoological legal model has no significance regarding
ai entities. For example, cutting off a dog’s leg for no medical reason is
considered abuse, and in some countries it is considered an offense. But
no country considers cutting off a robot’s arm an offense or abuse, even if
the robot is a humanoid, and even if the robot is programmed to respond

to this action “emotionally.”83
Therefore, given that the law prefers rationality over emotionality when
evaluating legal responsibility, and because the rationality of ai technology
is based on formal human logical reasoning, ai technology is much closer
to humans than to animals from the point of view of the law and with
regard to all legal aspects. So the zoological legal model is not suitable for
evaluating the legal responsibility of ai entities.84 To subject ai entities to
the criminal law, it is necessary to use the basic concepts of criminal li-
ability, which form the general requirements of criminal liability and must
be met in order to impose criminal liability on any individual (human,
corporate, or ai ) entity.
1.3. The Moder n Offe nder

Who is considered an offender under modern criminal law? Most people
associate the terms criminal and offender with “evil.” Murderers and rap-
ists are criminals; therefore, they are evil. Within the confines of criminal-
ity, however, we encounter not only murder and rape, but other behaviors
as well, which are not considered evil by most people. Some “white-
collar” crimes are not always considered evil, and many people regard
them as rather clever ploys or as succumbing to complex bureaucratic
systems. Similarly, most breaches of traffic laws are not regarded as evil,
nor are some of the offenses committed under the pressures imposed by
discharging one’s duty.
For example, consider the case of a surgeon who conducts a compli-
cated emergency surgery to save a patient’s life, and must work under great
time pressure, so that the operation fails despite the surgeon’s earnest ef-
forts, and the patient dies. The post mortem examination reveals that one
of the surgeon’s acts may be considered negligent, and she becomes crimi-
nally liable for negligent homicide. Do we consider her evil?
Indeed, even some of the classic crimes, such as murder, do not always
appear evil under certain circumstances. For example, a son witnesses the
daily suffering of his mother, who is dying of a terminal illness; she asks
him to disconnect her from the cpr machine to end her suffering. If the son
reluctantly agrees, he is criminally liable for murder.85 But is he evil? And
what if he refuses his mother’s request, in order to watch and presumably
revel in her suffering? He would not be criminally liable for any offense,
but would we not consider him evil?
Countless similar examples could be cited. Evil is not a measure of
criminal liability. Morality of any kind and criminal liability are vastly
different. At times they coincide, but such coincidence is not necessary
for the imposition of criminal liability. An offender is a person on whom
criminal liability has been imposed. When the legal requirements of a
given criminal offense are met by individual behavior, criminal liability
is imposed, and the individual is considered an offender. Evil may or may
not be involved. Thus, imposition of criminal liability requires the explo-
ration of its requirements.
Criminal law is considered the most efficient means of social control.
Society, as an abstract body or entity, controls its individual members
using a variety of measures (for example, moral, economic, or cultural),
but one of the most efficient is the law. And given that only criminal law
includes significant sanctions (imprisonment, fines, and even death), it is
considered the most efficient of the means used to control individuals; this
is commonly called legal social control. Imposition of criminal liability
is the application and implementation of legal social control.
Modern criminal liability is independent of morality of any kind, and
independent of evil. It is imposed in an organized, almost mathematical
way. Two cumulative types of requirements must be met for the impo-
sition of criminal liability: the first has to do with the law, the second
with the offender. Both types of requirements must be met in order to
impose criminal liability, and if they are met, no additional conditions
are required.
1.3.1. Basic R equirements for a G iven O ffense

The first type of requirement, placed on the law, includes four separate
requirements that every offense must meet, as defined by law. If the of-
fense fails to meet even one of these requirements, no court can impose
criminal liability on individuals for that offense. The four requirements
are (1) legality, (2) conduct, (3) culpability, and (4) personal liability. Each
of these four requirements represents a fundamental principle in criminal
law, that is, the principle of legality, the principle of conduct, the principle
of culpability, and the principle of personal liability.
Legality is a requirement for an offense to be considered legal (nullum
crimen sine lege). It is legality that forms the rules used to determine what
is criminally “right” and “wrong.” For an offense to be considered legal,
it must meet the following four cumulative conditions: (1) legitimate legal
source, (2) applicability in time, (3) applicability in place, and (4) legiti-
mate interpretation.86 Only when all four conditions are met is the offense
considered legal.
The offense must have a legitimate legal source that creates and defines
it.87 In most countries, the ultimate legitimate legal source is legislation,
and case law is not considered a legitimate source because legislation is
enacted by elected public representatives of society as a whole. Given that
criminal law exercises legal social control, it should reflect the will of
society by means of its representatives. The requirement that the offense
be applicable in time means that retroactive offenses are illegal.88 In plan-
ning their moves, individuals must know the prohibitions in advance, not
retroactively. Only in rare cases are retroactive offenses considered legal,
for example when a new offense is to the benefit of the defendant (it is
more lenient than its predecessor or allows for a new defense) or when the
offense covers cogent international custom, or jus cogens, (for example,
genocide, crimes against humanity, or war crimes).89
The offense must also be applicable in place, so that extraterritorial of-
fenses are illegal.90 Criminal law is based on the authority of the sovereign,
which is domestic, meaning that criminal law must be domestic as well.
For example, the criminal law of France is applicable in France but not
in the United States. Thus, extraterritorial offense is illegal (for example,
French offense that is applicable in the United States). In rare cases, how-
ever, the sovereign is authorized to protect itself or its inhabitants abroad
through extraterritorial offenses, as, for example, in the case of the foreign
terrorists who attacked the US Embassy in Kenya and who may be indicted
in the United States under US criminal law, although they have never been
to the United States.91 Extraterritorial offenses may also apply in cases of
international cooperation between states.
Offenses must be well formulated and phrased. They must be general
and address an unspecified public (for example, “Samuel Jackson is not
allowed to . . .” is not a legitimate offense).92 Offenses must be feasible,
so that legal social control is realistic (for example, “Whoever breathes
for more than five minutes uninterruptedly shall be guilty of . . .” is not a
legitimate offense).93 They must also be clear and precise, so that individu-
als will know exactly what they are allowed to do and what is prohibited
conduct.94 When all these conditions are met, the requirement of legality
is satisfied.
Conduct is a requirement that every offense must meet in order to
be considered legal (nullum crimen sine actu). Modern society prefers
freedom of thought and has no interest in punishing mere thoughts (cogi-
tationis poenam nemo patitur). Effective legal social control cannot be
achieved through mind policing, which is not enforceable. The conduct
is the objective and external expression of the commission of the offense.
Offenses that do not meet this are not legitimate. Throughout legal human
history, only tyrants and totalitarian regimes have used offenses that lack
conduct.
The conduct requirement of some offenses can be satisfied by inac-
tion. Offenses that criminalize the status of the individual rather than his
or her conduct are considered status offenses. For example, offenses that
punish the relatives of traitors simply because they are relatives, regard-
less of their conduct, are considered status offenses.95 So are offenses that
punish individuals of certain ethnic origin.96 Most developed countries
have abolished these offenses, and defendants indicted for such offenses
are acquitted in court because status offenses contradict the principle of
conduct in criminal law.97 Only when conduct is required may offenses be
considered legal and legitimate.
The culpability requirement (nullum crimen sine culpa) must be met
for an offense to be considered legal. Modern society has no interest in
punishing accidental, thoughtless, or random events, only events that are
the result of an individual’s culpability. Someone’s death does not nec-
essarily indicate the presence of an offender. A person may pass next to
another exactly when the other person is struck dead by lightning, but
the passerby is not necessarily culpable. The offense must require some
level of culpability for the imposition of criminal liability; otherwise, this
would amount to cruel maltreatment of individuals by society.
Culpability relates to the mental state of the offender and reflects the
subjective-internal expression of the commission of the offense. The re-
quired mental state of the offender, which forms the requirement of culpa-
bility, may be reflected both in the particular requirements of the specific
offense and in the general defenses. For example, the offense of man-
slaughter requires recklessness as its minimal level of culpability.98 But if
the offender is insane (general defense of insanity),99 is an infant (general
defense of infancy),100 or has acted in self-defense (general defense of self-
defense),101 no criminal liability is imposed. No culpability can be ascribed
to such an offender. An offense is considered legal and legitimate only if
it requires culpability.
Personal liability is required for an offense to be considered legal.102
Modern society has no interest in punishing one person for the behavior
of another, regardless of their relationship. Effective legal social control
cannot be achieved unless all individuals are liable for their own behavior.
If someone knows that no legal liability is imposed on him for his behav-
ior, he has no incentive to avoid committing offenses or other types of anti
social behavior. Legal social control can be effective only when a person
knows that for his own behavior, no other person is liable. Punishment
deters individuals only if they may be punished personally.
Personal liability guarantees that each offender would be criminally
liable and punished only for his or her own behavior. When individuals
collaborate in committing an offense, each of the accomplices is criminally
liable only for his or her own part. The accessory is criminally liable for
accessoryship, and the joint perpetrator for joint perpetration. The various
types of criminal liability in conjunction with the principle of personal
liability have established the general forms of complicity in criminal law
(for example, joint perpetration, perpetration-through-another, conspiracy,
incitement, and accessoryship). Only when personal liability is required
can the offense be considered legal and legitimate.
When all four basic requirements of legality, conduct, culpability, and
personal liability are met, the offense that embodies them is considered to
be legitimate and legal. Only then can society impose criminal liability on
individuals for the commission of these offenses. The legitimacy and legal-
ity of specific offenses is necessary for the imposition of criminal liability,
but not sufficient. For the imposition of criminal liability, the particular
requirements of the offense must be met by the offender as well. These
requirements are part of the definition of the offense.
1.3.2. R equirements of a G iven O ffense:

M odern C riminal L iability as a M atrix of M inimalism
The second type of requirement is focused on the offender’s behavior. Each
offense that meets the requirements to be considered such, determines
the requirements needed for the imposition of criminal liability based on
that offense. Although different offenses have differing requirements, the
formal logic behind all offenses and their structure is similar. The common
formal logic and structure are significant attributes of modern criminal li-
ability. In general, these attributes are characterized by minimalism, which
in this context means that the offense determines only the lower threshold
for the imposition of criminal liability. In other words, the offender is re-
quired to meet at least the requirements of the given offense. There are two
general requirements for an offense: (1) the factual element requirement
and (2) the mental element requirement.
The modern structure of the factual element requirement is the same
in most developed countries. This structure applies the fundamental prin-
ciple of conduct to criminal liability, and it is identical for every type
of offense, regardless of its mental element requirement. The factual ele-
ment requirement is the broad objective-external basis of criminal liability
(nullum crimen sine actu),103 designed to answer four main questions about
the factual aspects of the delinquent event: “What?” “Who?” “When?” and
“Where?” “What” refers to the substantive facts of the event (what hap-
pened). “Who” relates to the identity of the offender. “When” addresses
the time of the event. “Where” specifies the location of the event.
In some offenses, these questions are answered directly within the
definition of the offense. In others, some of the questions are answered
through the applicability of the principle of legality in criminal law.104
For example, the offense “Whoever assaults another person . . .” does not
relate directly to the questions of “who,” “when,” and “where,” but the
questions are answered through the applicability of the principle of legal-
ity. Because the offense is likely to be general,105 the answer to the “who”
question is any person who is legally competent. As this type of offense
may not be applicable retroactively,106 the answer to the “when” ques-
tion is from the time the offense was validated onward. And because this
type of offense may not be applicable extraterritorially (with some excep-
tions),107 the answer to the “where” question is within the territorial juris-
diction of the sovereign (with some exceptions).
Nevertheless, the answer to the “what” question must be incorporated
directly into the definition of the offense. This question addresses the core
of the offense and cannot be answered through the principle of legality.
This approach is at the foundation of the modern structure of the factual
element requirement, which consists of three main components: conduct,
circumstances, and results. Conduct is a mandatory component, whereas
circumstances and results are not. Thus, if the offense is defined as having
no conduct requirement, it is not legal, and the courts cannot convict indi-
viduals based on such a charge. Table 1.1 lists the components that answer
the four questions.
The conduct component is at the heart of the answer to the “what”
question. Status offenses, in which the conduct component is absent, are
considered illegal, and in general they are abolished when discovered.108
But the absence of circumstances or of results in the definition of an of-
fense does not invalidate the offense.109 These components are aimed at
meeting the factual element requirement with greater accuracy than by
conduct alone. Therefore, there are four formulas that can satisfy the fac-
tual element requirement:
1. conduct
2. conduct + circumstances
3. conduct + results
4. conduct + circumstances + results
Table 1.1 . Target Questions of the Factual Element Components
f ac tual elemen t componen ts que stion s
conduct “what”
circumstances “what,” “who,” “when,” “where”
results “what”
The modern structure of the mental element requirement has been

widely accepted in most developed countries. This structure applies the
fundamental principle of culpability in criminal law (nullum crimen sine
culpa). The principle of culpability has two main aspects: positive and
negative. The positive aspect (what needs to be in the offender’s mind in
order to impose criminal liability) relates to the mental element, whereas
the negative aspect (what should not be in the offender’s mind in order to
impose criminal liability) relates to the general defenses.110 For example,
imposition of criminal liability for manslaughter requires recklessness as
mental element, but it also requires that the offender not be insane. Reck-
lessness is part of the positive aspect of culpability, and the general de-
fense of insanity is part of its negative aspect.
The positive aspect of culpability in criminal law concerns the in-
volvement of mental processes in the commission of the offense. The
mental element has two important aspects: cognition and volition. Cogni-
tion is the individual’s awareness of the factual reality. In some countries,
awareness is called knowledge, but in this context there is no substantive
difference between awareness and knowledge, both of which can relate
to data from the present or the past, but not from the future.111 A person
may assess or predict what will be in the future, but not know or be aware
of it. Prophetic skills are not required for criminal liability. Cognition in
criminal law refers to a binary situation: the offender either is aware of
fact X or is not. Partial awareness is not accepted in criminal law, and it is
classified as unawareness.
Volition has to do with the individual’s will, and it is not subject to
factual reality. An individual may wish for unrealistic events to occur
or to have occurred, in the past, the present, or the future. Volition is
not binary because there are different levels of will, the three basic ones
being positive (A wants X), neutral (A is indifferent toward X), and nega-
tive (A does not want X). The cognitive and volitive aspects combine
to form the mental element requirement, as derived from the positive
aspect of culpability in criminal law. In most developed countries, there
are three principal forms of mental element, which are differentiated

based on the cognitive aspect. The three forms represent three layers of
positive culpability: (1) mens rea (general intent), (2) negligence, and
(3) strict liability.
The highest layer of the mental element is that of mens rea, which
requires full cognition. The offender is required to be fully aware of the
factual reality. This layer involves examination of the offender’s subjective
mind. Negligence, the second layer, is cognitive omission: the offender is
not required to be aware of the factual element, although based on objec-
tive characteristics, she could and should have had awareness of it. Strict
liability is the lowest layer of the mental element; it replaces what was
formerly known as absolute liability. Strict liability is a relative legal pre-
sumption of negligence based on the factual situation alone, which may be
refuted by the offender (praesumptio juris tantum). Cognition relates the
factual reality, as just noted. The relevant factual reality in criminal law is
that which is reflected by the components of the factual element.
From the perpetrator’s point of view, however, when committing the
offense, only the conduct and circumstance components of the factual
element exist in the present; the results occur in the future. Cognition is
restricted to the past and the present, and therefore it relates to conduct
and circumstances. Although results occur in the future, the possibility of
their occurrence ensuing from the relevant conduct exists in the present,
so that cognition can also relate to the possibility of the occurrence of the
results. For example, in the case of homicide, A aims a gun at B and pulls
the trigger. At this point, A is aware of his conduct, of the existing circum-
stances, and of the possibility of B’s death as a result of his conduct.
Volition is considered immaterial for both negligence and strict liabil-
ity, and can be added only to the mental element requirement of mens rea,
which embraces all three basic levels of will. Table 1.2 compares the gene
ral requirements of the three layers, or forms, of mental element.
Because in most legal systems, the default requirement for the mental
element is mens rea, negligence and strict liability offenses must specify
explicitly the relevant requirement. The explicit requirement may be listed
as part of the definition of the offense or included in the explicit legal tra-
dition of interpretation. If no explicit requirement of this type is men-
tioned, the offense is classified as a mens rea offense, which is the default
requirement. The relevant requirement may be met not only by the same
form of mental element, but also by a higher-level form. Thus, the mental
element requirement of the offense is the minimal level of mental element
needed to impose criminal liability;112 a lower level is insufficient for im-
Table 1.2 . Schematic Comparison of the Requirements

of Mental Element Forms
re quiremen t
Form Cognition Volition
Mens rea subjective cognition only in result offenses

(because of the relevant
level)
Negligence cognitive omission none
Strict liability presumed cognitive none
omission
Table 1.3 . Schematic Comparison of the Requirement

for Satisfying the Mental Element Forms
MEN TAL ELEMEN T RE QUIREMEN T OF . . . I S SA TI SFIE D THRO UGH . . .
mens rea negligence mens rea negligence or

strict liability mens rea strict liability or
negligence or
mens rea
posing criminal liability for the offense. Table 1.3 compares schematically
the requirements for satisfying the three forms of mental element.
This structure of criminal liability forms a matrix of minimalist re-
quirements. Each offense embodies the minimum requirement for the im-
position of criminal liability, and meeting these requirements is sufficient
to impose criminal liability. No additional psychological elements are re-
quired, so that the modern offender is an individual who meets the mini-
mal requirement of an offense. The offender is not required to be “evil,”
“mean,” or “wicked” — only to meet all the requirements of the offense.
The imposition of criminal liability is therefore dry and rational (as op-
posed to emotional), and resembles mathematics.
Thus, the matrix of minimalist requirements for criminal liability has
two aspects: structural and substantive. For example, if the mental ele-
ment of the offense requires only awareness, then no other component of
the mental element is required (structural aspect), and the required aware-
ness is defined by criminal law irrespective of its meaning in psychol-
ogy, philosophy, theology, and so on (substantive aspect). The substantive
aspect is discussed in detail in chapters 2, 3, and 4.
Nevertheless, it is clear that the structure of criminal liability has

been designed for humans, with the abilities of humans — and not other
creatures — in mind. The mental element requirement relies on the human
spirit, soul, and mind. Can machines, therefore, be examined based on
human standards of spirit, soul, and mind? Considering these insights,
how can criminal liability be imposed on entities lacking in spirit and
soul?
It is important to remember, however, that although insights in crimi-
nal liability rely on the human spirit and soul, the imposition of criminal
liability itself does not depend on these terms that carry deep psychologi-
cal meaning. If an entity, any entity, meets both the factual and mental ele-
ment requirements of the offense, then criminal liability may be imposed
with or without spirit, with or without soul. This understanding is not
new, and perhaps it is not highly innovative in the twenty-first century.
A similar understanding was reached in the seventeenth century, long
before the emergence of modern ai technology, but at a time when other
nonhuman creatures committed offenses and it was necessary to subject
them to criminal law. These creatures have neither spirit nor soul, but they
have the ability to be subjects of criminal liability, which is being imposed
on them to this day. These creatures are corporations.
1.3.3. The C ase of N onhuman C orporations (R ound 1)

Since the seventeenth century, corporations have been among the modern
offenders, although they lack spirit, soul, and a physical body.113 Corpo-
rations were recognized already by Roman law, but the evolution of the
modern corporation began in the fourteenth century, when English law
demanded permission from the king or Parliament to recognize a cor-
poration as legal.114 Early corporations in the Middle Ages were mostly
ecclesiastical bodies that were active in the organization of church prop-
erty. From these bodies evolved associations, commercial guilds, and
professional guilds that eventually formed the basis for commercial
corporations.
During the sixteenth and seventeenth centuries, hospitals and univer-
sities were also commonly incorporated.115 In addition to these, commer-
cial corporations developed to solve problems of division of ownership
among several owners of a business.116 When people created a new busi-
ness, ownership could be divided between them by establishing a corpo-
ration and dividing the shares and stocks among the shareholders. This
pattern of ownership division has been perceived as efficient in minimiz-
ing the risks of the owners regarding the financial hazards of the business,
and as a result, corporations have become common.117
The use of corporations flourished during the first industrial revolu-
tion, to the point where they have been identified both with the fruits of
the revolution and with the misery of the lower classes and the workers
(who were created by the revolution). Corporations were regarded as being
responsible for the poverty of the workers, who did not share in the profits,
and for the continuing abuse of children employed by the corporations.
As the industrial revolution progressed, public and social pressure on the
corporations increased, until legislators were forced to restrict the activi-
ties of corporations.
In the beginning of the eighteenth century, the British Parliament en-
acted statutes against the abuse of power by corporations, the very power
granted to them by the state for the benefit of social welfare.118 To ensure
the effectiveness of these statutes, they included criminal offenses. The
offense most often used for this purpose was that of public nuisance.119
This legislative trend deepened as the revolution progressed, and in the
nineteenth century, most developed countries had advanced legislation
regarding the activities of corporations in a variety of contexts. To be
effective, this legislation included criminal offenses as well, prompt-
ing the conceptual question, how can criminal liability be imposed on
corporations?
Criminal liability requires a factual element, whereas corporations pos-
sess no physical body. Criminal liability also requires a mental element,
whereas corporations have no mind, brain, spirit, or soul.120 Some Euro-
pean countries refused to impose criminal liability on nonhuman crea-
tures, and revived the Roman rule whereby corporations were not subject
to criminal liability (societas delinquere non potest). But this approach
was highly problematic because it created legal shelters for offenders. For
example, an individual who did not pay taxes was criminally liable, but
when the individual was a corporation, it was exempt. This provided an
incentive for individuals to work through corporations and evade paying
taxes. Eventually, all countries subjected corporations to criminal law, but
not until the twentieth century.
The Anglo-American legal tradition accepted the idea of the criminal
liability of corporations early because of its many social advantages and
benefits. In 1635, criminal liability for corporations was enacted and im-
posed for the first time.121 This was a relatively primitive imposition of
criminal liability because it relied on vicarious liability, but it enabled
the courts to impose criminal liability on corporations in a way that was

separate from the criminal liability of any of its owners, workers, or share-
holders. This structure remained in effect throughout the eighteenth and
nineteenth centuries.122
The major disadvantage of criminal liability based on vicarious liabil-
ity was that it required valid vicarious relations between the corporation
and another entity, which in most cases happened to be human.123 Con-
sequently, when the human entity acted without permission (ultra vires),
the corporation was exempt of responsibility. To exempt a corporation of
liability, it was sufficient to include a general provision in its incorpora-
tion papers prohibiting the commission of any criminal offense on behalf
of the corporation.124 This deficiency resulted in the replacement in Anglo-
American legal systems of the model of criminal liability for corporations
during the late nineteenth century and early twentieth century.125
The new model was based on the identity theory. In some types of cases,
the criminal liability of corporations derives from its organs, whereas in
other types of cases, criminal liability is independent. When the criminal
offense requires an omission (for example, not paying taxes, not fulfilling
legal requirements, or not observing workers’ rights), and the duty to act
is the corporation’s, the corporation is criminally liable independently, re-
gardless of any criminal liability of other entities, whether human or not.
When the criminal offense requires an act, the acts of its organs are related
to the corporation if they were committed on its behalf, with or without
permission.126 The same structure works for the mental element, for mens
rea, negligence, and strict liability alike.127
Consequently, the criminal liability of the corporation is direct, not
vicarious or indirect.128 If all requirements of an offense are met by the
corporation, it is indicted regardless of any proceedings that may be in
effect against any human entity. If convicted, the corporation is punished
separately from any human entity. Punishments for corporations are con-
sidered to be not less effective than for humans. To complete this picture,
we will return to discuss the case of nonhuman corporations (Round 2) in
relation to the question of the sentencing of ai entities.129
However, the main significance of the modern legal structure of crimi-
nal liability of corporations is conceptual. Since the seventeenth century,
criminal liability has not been a uniquely human domain. Other entities,
nonhuman, are also subject to the criminal law. Some adjustments were
needed to make this legal structure applicable, but eventually, nonhuman
corporations have become subject to criminal law. Today, this seems en-
tirely natural, as it should. Given that the first barrier was crossed in the
seventeenth century, the road may be open to crossing the next barrier that
stands before the imposition of criminal liability on ai entities.
In the following chapters, the criminal liability of ai entities is ana-
lyzed in detail for all types of offenses. Because the major obstacle on the
way to criminal liability for ai entities is that of the mental element, the of-
fenses are divided according to their mental element requirements (mens
rea, negligence, and strict liability). Next, the applicability of general de-
fenses for ai entities is analyzed. Finally, in the last chapter, the applicabil-
ity of human punishments to ai entities is discussed.
2
ai C riminal L ia bili ty for I n ten tional O ffen se s
In 1981, a thirty-seven-year-old Japanese employee in a motorcycle factory

was killed by an ai robot working near him. The robot identified the em-
ployee as a threat to its mission and calculated that the most efficient way
to eliminate the threat was to push the worker into an adjacent machine.
Using its very powerful hydraulic arm, the robot smashed the surprised
worker into the operating machine, killing him instantly, after which it re-
sumed its duties without further interference.1 The legal question is, who
is to be held criminally liable for this homicide?
Intentional offenses make up most of the criminal offenses in most
countries. To impose criminal liability in these cases, both factual element
and mental element requirements must be met by the apparent offender.
Consequently, the question of the criminal liability of ai entities for in-
tentional offenses is whether ai entities have the capability to meet these
requirements accumulatively. Note, however, that even if ai entities have
such capability, this does not exempt humans from their criminal liability,
if humans are involved in the commission of the offense.
To explore these issues, we examine the capability of ai entities to
meet the requirements of both factual and mental elements of intentional
offenses to allow the imposition of human criminal liability for intentional
ai offenses.
2.1. The F ac tual Element R equire ment

The factual element requirement structure (actus reus) is identical for all
types of offenses: intentional, negligence, and strict liability, as previously
noted.2 This structure contains one mandatory component (conduct) for
all offenses, and two optional components (circumstances and results) for
some offenses. Let us now discuss the capability of machines to meet the
factual element requirement considering this structure.
38
AI C r iminal L iabili ty for Intention al Offen ses . 39
2.1.1. C onduct
Conduct, as already noted, is the only mandatory component of the factual
element requirement; in other words, an offense that does not require con-
duct is illegal and illegitimate, and no criminal liability can be legitimately
imposed based on it. In independent offenses, the conduct component of
the factual element requirement may be expressed both by an act and by
omission. In derivative criminal liability, the conduct component may be
expressed also by inaction, with some restrictions. We will examine these
forms of conduct in relation to the capabilities of ai machines.
In criminal law, act is defined as material performance with a factual-
external presentation. According to this definition, the materiality of the
act is manifest in its factual-external presentation, which differentiates
the act from subjective-internal matters related to the mental element.
Because thoughts have no factual-external presentation, they are related
not to the factual, but to the mental element. Will may initiate acts, but
in itself has no factual-external presentation, and is therefore considered
part of the mental element. Consequently, involuntary and unwilled ac-
tions, as well as reflexes, are still considered acts as far as the factual ele-
ment requirement is concerned.3
But although unwilled acts or reflexes are considered acts, criminal
liability is not necessarily imposed for such offenses because of reasons
involving the mental element requirement or general defenses. For ex-
ample, B physically pushes A in the direction of C. Although the physical
contact between A and C is involuntary for both, it is still considered an
act. But it is unlikely that criminal liability for assault will be imposed on
A, because the mental element requirement for assault has not been met.
Acts committed as a result of loss of self-control are still considered acts,
but loss of self-control is a general defense, which if proven and accepted,
exempts the offender from criminal liability.
This definition of an act concentrates on its factual aspects to the ex-
clusion of mental aspects.4 The definition is also sufficiently broad to in-
clude actions originating in telekinesis, psychokinesis, and so on, if these
are possible,5 as long as they have a factual-external presentation.6 If act
were restricted to “willed muscular construction” or “willed bodily move-
ment,”7 this would prevent imposition of criminal liability in cases of
perpetration-through-another (for example, B in the example just cited)
because no act has been performed, and it would be possible to assault
someone and be exempt from criminal liability, by pushing an innocent
person upon the victim.
Consequently, criminal law considers as an act any material perfor-

mance with a factual-external presentation, whether willed or not. On
this basis, ai technology is capable of performing acts that satisfy the con-
duct requirement. This is true not only for strong ai technology, but also
for much lower technologies. When a robot moves its arms or any other
of its devices, it is considered to act. This is true when the movement is
the result of inner calculations carried out by the robot, but not only then.
Even if the robot is fully actuated by a human operator through remote
control, any movement of the robot is considered an act.
Therefore, even non – a i technology robots have the factual capabil-
ity of performing acts, regardless of the motives or reasons for the act.
This does not necessarily mean that these robots are criminally liable for
their acts, because the imposition of criminal liability must also satisfy
the mental element requirement. But as far as performing an act in order
to satisfy the conduct component requirement is concerned, any mate-
rial performance with a factual-external presentation is considered an act,
whether the physical performer is a strong ai entity or not.
In criminal law, omission is defined as inaction contradicting a legiti-
mate duty to act. According to this definition, the term legitimate duty is
of great significance. The opposite of action is not omission, but inaction.
If doing something is an act, then not doing anything is inaction. Omission
is an intermediate degree of conduct between action and inaction. Omis-
sion is not mere inaction, but is inaction that contradicts a legitimate duty
to act: the omitting offender is required to act, but fails to do so.8 If no act
has been committed, but no duty to act is imposed, no omission has been
committed.9
The omitting offender is, therefore, punished for doing nothing in spe-
cific situations where he or she should have done something owing to a
certain legitimate duty. For example, in most countries parents have a legal
duty to care for their children. In these countries, the breach of this duty
may form an offense. In this situation, the parent is not punished for acting
in a wrong way, but for not acting although having a legal duty to act in a
specific way. The requirement to act must be legitimate in the given legal
system, and in most legal systems, the legitimate duty may be imposed
both by law (legal duty) and by contract (contractual duty).10
The modern concept of conduct in criminal law acknowledges no sub-
stantive or functional differences between acts and omissions for the im-
position of criminal liability.11 Therefore, any offense may be committed
both by act and by omission. Socially and legally, commission of offenses
by omission is no less severe than commission by act.12 Most legal systems
accept this modern approach, and there is no need to explicitly require

omission to be part of the factual element of the offense. The offense de-
fines the prohibited conduct, which may be committed both through acts
and through omissions.13
On this basis, ai technology can perform omissions that satisfy the
conduct requirement. This is true not only for strong ai technology, but
also for much lower technologies. Physically, commission by omission
requires doing nothing. There is no doubt that any machine is capable of
doing nothing; therefore, any machine is physically capable of committing
an omission. Naturally, for the inaction to be considered omission, there
must be a legal duty that contradicts the inaction. If such duty exists, origi-
nating in law or contract, and the duty is addressed to the machine, then
there is no question that the machine is capable of committing an omission
with respect to that duty.
This is also the situation regarding inaction. Inaction is the complete
factual opposite of an act. If an act is doing something, then inaction
is not doing it, or doing nothing. Whereas omission is inaction contra-
dicting a legitimate duty to act, inaction requires no such contradiction.
Omission is not doing when there is an obligation to do, whereas inac-
tion is not doing when there is no obligation to do anything. Inaction
is accepted as a legitimate form of conduct only in derivative criminal
liability (for example, attempt, joint perpetration, perpetration-through-
another, incitement, or accessoryship), and not in complete and inde-
pendent offenses.14
In these instances, when inaction is accepted as a legitimate form of
conduct, it is physically committed in the same way as omission. Con-
sequently, if ai technology is capable of commission through omission
of conduct, it is also capable of commission through inaction. This does
not necessarily mean that machines or robots are automatically criminally
liable for these omissions or inactions, because to impose criminal liabil-
ity, it is necessary to also meet the mental element requirement, and it is
not sufficient to satisfy the factual element requirement.
Thus, the mandatory component of the factual element requirement
(conduct) can be satisfied by machines. These machines need not be
highly sophisticated or even based on ai technology. Simple machines are
capable of conduct under the definitions and requirements of criminal law.
For the imposition of criminal liability on any entity, this is an essential
step, even if not a sufficient one. No criminal liability can be imposed if
the conduct requirement is not satisfied, but conduct alone is not sufficient
for the imposition of criminal liability.
2.1.2. C ircumstances
Circumstances are not a mandatory component of the factual element re-
quirement. Some offenses require circumstances in addition to conduct,
and some do not. Circumstances are defined as factual components that
describe the conduct but do not derive from it. According to this defini-
tion, circumstances specify the criminal conduct in more accurate terms.
When defining specific offenses, circumstances are required especially
when the conduct component is too broad or vague, and greater specific-
ity is needed in order to avoid over-criminalization of situations that are
considered legal by society.
In the case of most offenses, the circumstances represent the factual
data that make the conduct criminal. For example, in most legal systems,
the conduct in the offense of rape is having sexual intercourse (although
the specific verbs may vary). But in itself, having sexual intercourse is not
an offense, and it becomes one only if it is committed without consent.
The factual element of “without consent” is what makes the conduct of
“having sexual intercourse” criminal. In this offense, the factual compo-
nent “without consent” functions as a circumstance.15 Additionally, the
factual component “with a woman” also functions as circumstance be-
cause it describes the conduct (raping a chair is not an offense).
In this example, the factual element of the offense of rape is “having
sexual intercourse with a woman without consent.”16 Whereas “having
sexual intercourse” is the mandatory conduct component, “with a woman”
and “without consent” function as circumstance components that de-
scribe the conduct more accurately, to make sure that the factual element
requirement of rape is specified adequately, in a way that avoids over-
criminalization.
According to the definition, circumstances are not derived from the
conduct in order to allow distinguishing them from the result compo-
nent.17 For example, in homicide offenses, the conduct is required to cause
the “death” of a “human being.” Death describes the conduct and also de-
rives from it because it is the conduct that caused death. Therefore, in ho-
micide offenses, “death” does not function as a circumstance. The factual
data that functions as a circumstance is “human being.” The victim was a
human being long before the conduct took place, and therefore it does not
derive from the conduct. But these words also describe the conduct (caus-
ing the death of a human being, not of an insect), and therefore function
as a circumstance.
On this basis, ai technology is capable of satisfying the circumstance
component requirement of the factual element. When the circumstances
are external to the offender, the identity of the offender is immaterial, and
therefore the offender may be human or a machine, without affecting the
circumstances. In the earlier example of rape, the circumstance “with a
woman” is external to the offender. The victim is required to be a woman
regardless of the identity of the rapist. The raped woman is still considered
“woman” whether she was attacked by a human or a machine, or has not
been attacked at all.
When the circumstances are not external to the offender but are re-
lated somehow to the offender’s conduct, the circumstances assimilate
within the conduct component. In the example of rape, the circumstance
“without consent” describes the conduct as if it were part of the conduct
(how exactly did the offender have sexual intercourse with the victim).
To satisfy the requirement of this type of circumstance, the offender must
commit the conduct in a particular way. Consequently, meeting the re-
quirement of this type of circumstance is the same as committing the con-
duct. As already discussed, it is within the capabilities of machines to
meet the conduct component requirement.
2.1.3. R esults and C ausal C onnection

Results are not a mandatory component of the factual element require-
ment. Some offenses require results in addition to conduct, and some do
not. Results are defined as a factual component that derives from the con-
duct. According to this definition, results specify the criminal conduct
with greater accuracy. Results are defined as deriving from the conduct in
order to allow distinguishing them from the circumstances. For example,
in homicide offenses, the conduct is required to cause the “death” of a
“human being.” The death describes the conduct and also derives from it,
because the conduct caused the death. Therefore, in homicide offenses,
“death” does not function as a circumstance, but as a result.18
In the context of the factual element, the results derive from the con-
duct through factual causal connection. Although additional conditions
exist for this derivation (legal causal connection), the factual requirement
is a factual causal connection. Consequently, proof of results requires
proving a factual causal connection.19 The factual causal connection is
defined as a derivative connection, in which, were it not for the conduct,
the results would not have occurred the way they have. According to this
definition, the results are the ultimate consequences of the conduct, that
is, causa sine qua non.20
The factual causal connection relates not only to the mere occurrence
of the results, but also to the way in which they occurred. For example, A
hit B with her car, and B died. Because B was terminally ill, A may argue
that B would have died anyway in the near future, so B’s death is not the
ultimate result of A’s conduct, and the result would have occurred anyway,
even without A’s conduct. But because the factual causal connection has
to do with the way in which the results occurred, the requirement is met
in this example: B would not have died the way he did had A not run
him over.
On this basis, ai technology is capable of satisfying the result com-
ponent requirement of the factual element. To achieve the results, the of-
fender must initiate the conduct. Commission of the conduct forms the
results, and the results are examined objectively to determine whether
they are derived from the commission of the conduct.21 Thus, when the
offender commits the conduct, and the conduct has been carried out, it is
the conduct, not the offender, that is the cause of the results, if there was
a result. The offender is not required to cause, separately, any results, but
only to commit the conduct. Although the offender begins the factual pro-
cess that forms the results, this process is initiated only through the com-
mission of the conduct component.22
Because ai technology is capable of committing conducts of all types,
in the context of criminal law it is capable of causing results deriving from
this conduct. For example, when a robot operates a firearm and causes it
to shoot a bullet toward a human individual, this fulfills the conduct com-
ponent requirement of homicide offenses. At that point, a causal connec-
tion test is used to examine the conduct and determine whether it caused
that individual’s death. If it did, both the result and conduct component
requirements are met, although physically the robot “did” nothing but
commit the conduct. Because meeting the conduct component require-
ment is within the capabilities of machines, as already discussed, so is
meeting the results component requirements.
2.2. The Mental Element R equire ment

The imposition of criminal liability for intentional offenses requires the
fulfillment of both factual and mental element requirements. The mental
element requirement of intentional offenses is general intent, more accu-
rately known as mens rea, meaning “evil mind” in Latin. Because the term
general intent can at times be confused with intent or specific intent, we
prefer to use the more accurate term, mens rea. If ai technology is capable
of fulfilling the mens rea requirement, it is possible, feasible, and achiev-
able to impose criminal liability on ai entities for intentional offenses.
First, we explore the structure of the mens rea requirement, then we turn
to the fulfillment of the requirement by ai technology.
2.2.1. Structure of the M ens R ea R equirement

In the modern age, mens rea expresses basic types of mental elements, as
it embodies the idea of culpability most effectively. This is the only mental
element that combines cognition and volition. The mens rea requirement
expresses the internal-subjective relation of the offender with the factual
commission of the offense.23 In most legal systems, the mens rea require-
ment functions as the default option of the mental element requirement.
Therefore, unless negligence or strict liability are explicitly required as
mental elements of the offense, mens rea is the required mental element.
This default option is also known as the presumption of mens rea.24
Accordingly, all offenses are presumed to require mens rea, unless there
is some explicit deviation from this requirement. Because mens rea is the
highest level of known mental element requirement, this presumption has
great significance. Indeed, most offenses in criminal law require mens rea
and not negligence or strict liability. None of the mental element compo-
nents, including mens rea components, is independent or stands alone for
itself. For example, the dominant component of mens rea is awareness. If
we require the offender to be aware, the question arises, aware of what?
Awareness cannot stand alone because it would be meaningless.
Consequently, all mental element components must relate to facts or
to some factual reality. The relevant factual aspect for criminal liability
consists, naturally, of the factual element components (conduct, circum-
stances, and results), as discussed earlier.25 Factual reality contains many
more facts than the components of factual element, but all other facts are
irrelevant for the imposition of criminal liability. For instance, in the rape
example, the relevant facts are “having sexual intercourse with a woman
without consent,”26 and the rapist must be aware of these facts. Whether
the offender was or was not aware of other facts as well (for example, the
color of the woman’s eyes, the fact that she was pregnant, or her suffering)
is immaterial for the imposition of criminal liability.
Thus, for the imposition of criminal liability, the object of the mental
element requirement is the factual element components. Naturally, this
object is much narrower than the complete factual reality, but the factual
element represents the decision of society regarding what is relevant for
criminal liability and what is not. Nevertheless, other facts and the mental
relation to them may affect the punishment, but they are immaterial for
the imposition of criminal liability. For example, a particularly cruel rapist
would be convicted of rape, whether he was cruel or not, but his punish-
ment would likely be much harsher than that of a less cruel rapist.
Identifying the factual element components as the object of mens rea
components is the basis for the structure of mens rea. Mens rea has two
layers of requirement: cognition and volition. The layer of cognition con-
sists of awareness. The term knowledge is also used to describe the cogni-
tion layer, but awareness seems more accurate. Both awareness and knowl-
edge function the same way and have the same meaning in this context.
We can be aware only of facts that occurred in the past or are occurring at
present, but we cannot be aware of future facts.
For example, we can be aware of the fact that A ate an ice-cream cone
two minutes ago, and we can be aware of the fact that B is eating one right
now. If C said that she intends to eat an ice-cream cone, we can predict it,
foresee it, or estimate the probability that it will happen, but we cannot
be aware of it, simply because it has not occurred yet. If the criminal law
had required offenders to be aware of future facts, it would have required
prophetic skills. As far as we know, most criminals do not possess such
skills and are far from being prophets.
In this context, the offender’s temporal point of view determines the
point at which the conduct is actually performed. Therefore, from the of-
fender’s point of view, the conduct component always occurs in the pres-
ent. Consequently, awareness is a relevant component of mens rea in re-
lation to conduct. Circumstance is defined as factual data that describes
the conduct but does not derive from it.27 To describe current conduct,
circumstances must exist in the present as well. For example, the circum-
stance “with a woman” in the offense of rape, described earlier, must exist
simultaneously with the conduct “having sexual intercourse.”
The raped person must be a woman during the commission of the of-
fense for the circumstance requirement to be fulfilled. Therefore, aware-
ness is a relevant component of mens rea in relation to circumstances as
well. The situation is different, however, in relation to results, which are
defined as a factual component that derives from the conduct; this means
that results must occur after the conduct.28 For example, B dies at 10:00:00
after which, at 10:00:10, A shoots her. In this case, the conduct (the shoot-
ing) is not the cause of the other factual event (B’s death), which does not
function as a result.
From the offender’s point of view, given that the results occur after the
conduct, and because the offender’s temporal point of view determines
the point at which the conduct is performed, the results occur in the
future. Consequently, because results do not occur in the present, aware-
ness is not relevant for them. The offender is not expected to be aware of
the results, which have still not occurred from his point of view. But al-
though the offender cannot be aware of the future results, he is capable of
predicting them and assessing the probability of their occurrence. These
capabilities are present simultaneously with the actual performance of
the conduct.
For example, A shoots B. At the point in time when the shot is fired,
B’s death has not yet occurred, but in the act of shooting, A is aware of the
possibility of B’s death as a result of the shot. Therefore, awareness is a rel-
evant component of mens rea in relation to the possibility that the results
will occur, but not in relation to the results themselves. Awareness of this
possibility is not required to be related to how reasonable or probable the
occurrence of the result is. If the offender is aware of the mere existence of
the possibility, whether its probability is high or low, that the result will
follow from the conduct, this component of mens rea is present.29
The additional layer of mens rea is that of volition. This layer is ad-
ditional to cognition and based on it. Volition is never alone, but always
accompanies awareness. Volition relates to the offender’s will regarding
the results of the factual event. In rare offenses, volition may relate to mo-
tives and purposes beyond the specific factual event, and is expressed as
specific intent.30 The main question regarding volition (apart from the of-
fender’s awareness of the possibility that the results will follow from the
conduct) is whether the offender wanted the results to occur. Because from
the offender’s point of view the results occur in the future, they are the
only reasonable object of volition.
From the offender’s temporal point of view, neither circumstances nor
conduct have anything to do with will. The raped person was a woman
before, during, and after the rape, regardless of the rapist’s will. The sexual
intercourse is what it is at that point in time, regardless of the rapist’s will.
If the offender argues that the conduct occurred against his will — that is,
the offender did not control it — this argument is related to the general de-
fense of loss of self-control, which we will discuss in a moment.31 Thus, as
far as conduct and circumstances are concerned, no volition component is
required, only awareness.
In the factual reality there are many levels of will, but criminal law
accepts only three: (1) intent (and specific intent), (2) indifference, and (3)
rashness. The first represents positive will (the offender wants the results
to occur), the second represents nullity (the offender is indifferent about
the results), and the third represents negative will (the offender does not
want the results to occur, but has taken unreasonable risks that caused
them to occur). For example, in homicide offenses, at the moment when

the conduct is committed, if the offender
1. wants the victim to die, it is intent (or specific intent);

2. is indifferent as to whether the victim lives or dies, it is indif-
ference;
3. does not want the victim to die, but assumes unreasonable risk
in this regard, it is rashness.
Intent is the highest level of will accepted by criminal law. In most

countries, intended homicide is considered murder. Indifference is the
intermediate level, and rashness is the lowest level of will. Both indif-
ference and rashness are known as recklessness. In most countries, reck-
less homicide is considered manslaughter. Consequently, if the offense
requires recklessness, this requirement can be fulfilled by proof of intent
because a higher level of will satisfies lower levels. But if the offense re-
quires intent or specific intent, this requirement can be fulfilled only by
intent or specific intent.
It is easier to sum up the structure of mens rea if we divide offenses
into conduct offenses and result offenses. Conduct offenses are those
whose factual element requires no results; result offenses are those whose
factual element requires results.32 This division clarifies the structure of
mens rea because volition is required only in relation to results. Therefore,
results require both cognition and volition, whereas conduct and circum-
stances require only cognition. The structure of mens rea can be described
schematically as shown in table 2.1.
Therefore, in conduct offenses, whose factual element requirement
contains conduct and circumstances, the mens rea requirement contains
awareness of these components. In result offenses, whose factual element
requirement contains conduct, circumstances, and results, the mens rea
requirement contains awareness of the conduct, of the circumstances, and
of the possibility that the results will occur. In addition, in relation to re-
sults, the mens rea requirement contains intent or recklessness, according
to the definition of the offense.
This general structure of mens rea is a template that contains terms de-
rived from the terminology of the sciences of the mind (awareness, intent,
recklessness, and so on). To examine whether ai entities are capable of ful-
filling the mens rea requirement of given offenses, the definition of these
terms must be explored. Next we will discuss separately the cognitive and
volitive aspects of mens rea.
Table 2.1 . General Structure of Mens Rea
men s rea re quiremen t

type of offen se
Cognition Volition
Conduct offense awareness of conduct —

+
awareness of circumstances intent
awareness of conduct
+
Result offense awareness of circumstances or
+
awareness of the possibility recklessness
of the occurrence of results (indifference/
rashness)
2.2.2. F ulfillment of the C ognitive A spect

The cognitive aspect of mens rea contains awareness, and therefore the
relevant question is whether ai technology is capable of consolidating
awareness. Because the term awareness may have different meanings in
different scientific spheres (for example, psychology, theology, or law),
to answer this question we must examine the legal meaning of this term;
specifically, we must examine awareness based on its legal definition in
criminal law. Even if the legal meaning may differ from other meanings
of the term, given that our discussion relates to criminal liability, only the
legal meaning is relevant.
In criminal law, awareness is defined as perception of factual data by
the senses, and its understanding.33 The roots of this definition lie in the
psychological understandings of the late nineteenth century. Previously,
awareness was equated with consciousness, that is, the physiological-
bodily state the mind is in when the human is awake. By the end of the
nineteenth century, psychologists argued that awareness involves a much
more complicated situation of the human mind.34 In the 1960s, the modern
understanding of the human mind began to consolidate toward the con-
cept of the sum total of internal and external stimulations that the indi-
vidual is aware of at a given point in time.
Consequently, it became apparent that the human mind is not constant
but dynamic and regularly changing, and could be described as a flow of
feelings, thoughts, and emotions (“stream of consciousness”).35 It was also
understood that the human mind is selective, meaning that humans are ca-
pable of focusing their minds on certain stimuli while ignoring others, so
that the ignored stimuli do not enter the human mind at all. If the human
mind were to include all internal and external stimuli and had to pay at-
tention to each one of them, it would be incapable of functioning.
The function of the human and animal sensory systems is to absorb
the stimuli (light, sound, heat, pressure, and so on) and transfer them to
the brain for the processing of factual information. Processing is executed
in the brain internally until a relevant general image of the factual data is
created. This process is called perception, which is considered one of the
basic skills of the human mind.
Many stimuli are active at any one time. To enable the creation of an
organized image of factual data, the human brain must focus on some of
the stimuli and ignore others. This is accomplished through the process of
attention, which permits the brain to concentrate on some stimuli while
ignoring others. In practice, the other stimuli are not entirely ignored,
but they exist in the background of the perception process. The nervous
system still retains adequate vigilance that allows it to absorb other stimuli
when the process of attention is at work.
For example, Archie is watching television and is focused on it. Edith
calls him because she wants to speak with him. When she first calls, he
does not react. When she calls him for the second time, much louder, he
reacts, asks her what she wants, and says that he must go to the bathroom.
When Archie is focused on the television, many of the stimuli in his envi-
ronment are ignored (for example, the sound of his heartbeats, the smells
from the kitchen, and the pressure in his bladder), allowing him to focus
on the television. But his nervous system is still sufficiently alert to absorb
some of these stimuli. Edith’s first call is just another sound to be ignored
by the process of attention. But when she calls the second time, the at-
tention process used to focus on the television is stopped, and the other
stimuli are absorbed and receive some attention. This is why Archie sud-
denly remembers that he must go to bathroom.
Perception includes not only absorbing stimuli, but also processing
them into a relevant general image, which is what usually gives meaning
to the accumulation of stimuli. Processing the factual data into a rele
vant general image is accomplished through unconscious inference, so
that there is absolutely no need for awareness to carry this out.36 But the
result of the processing — that is, the relevant general image — is a con-
scious result. Thus, although the human mind is not conscious of most of
the process, it is conscious of its results when the relevant general image
is accepted. Consequently, the human mind is considered to be aware only

when the relevant general image is accepted.
This process forms the essence of human awareness, which is the final
stage of perception. Perception of factual data by the senses, and its un-
derstanding, results in the creation of a relevant general image, which is
the awareness of the factual data. For example, it is not the eyes but the
brain that is the sight organ of humans. Eyes function as mere sensors that
deliver factual data to the brain. Only when the brain creates the relevant
general image is the human considered to be aware of the relevant sight.
Consequently, a human in a vegetative situation, with fully functioning
eyes, is not considered to be seeing (or to be aware of what she sees),
unless the sights are combined into a relevant general image.
As a result, for us to consider a human to be aware of certain factual
data, two accumulative conditions are required: (1) absorbing the factual
data by senses, and (2) creating a relevant general image about these data
in the brain. If one of these conditions is missing, the person is not consid-
ered to be aware. Awareness is a binary issue: the person either is aware or
is not. Partial awareness is meaningless. An offender may be aware of part
of the factual data, that is, fully aware of some of the data, but not partly
aware of a certain fact. If the facts were not absorbed, or if no relevant
general image has been created, the offender is considered to be unaware.
At times, the term knowledge is used to describe the cognitive aspect
of mens rea, as already noted. Is there a difference between knowledge
and awareness? When examined functionally, there is no difference be-
tween these terms as far as criminal liability is concerned because they
both refer to the same idea of cognition.37 Moreover, knowledge has been
explicitly defined as awareness at times.38 But it seems that within the
context of the mental element, the more accurate term is awareness rather
than knowledge.
Outside of the criminal law context, awareness is more related to con-
sciousness than is knowledge. Knowledge is also related to the cognitive
process, but it refers more to information than consciousness does. Knowl-
edge represents perhaps a deeper cognitive process than awareness, but
within the specific context of the mental element requirement in criminal
law, awareness seems to be the more accurate term.
In court, proving the full awareness of the offender beyond any rea-
sonable doubt, as required in criminal law, is not an easy task. Awareness
relates to internal processes of the mind, which do not necessarily have
external expressions. Therefore, criminal law has developed evidential
substitutes for this task. These substitutes are presumptions, which in cer-
tain types of situations presume the existence of awareness. Two major

presumptions are recognized in most legal systems: (1) the willful blind-
ness presumption, as a substitute for awareness of conduct and of circum-
stances; and (2) the awareness presumption, as a substitute for awareness
of the possibility of the occurrence of the results. Before exploring these
presumptions and their relevance to ai technology, we must examine how
ai technology meets the cognitive aspect of the mens rea requirement.
Is an ai robot capable of being aware of conduct, circumstances, or the
possibility of the occurrence of the results in the context of criminal law?39
The process of awareness may be divided into two stages, as noted. The
first stage consists of absorbing the factual data by means of the senses.
At this stage, the primary role is played by the devices used to absorb the
factual data. The human devices are organs, such as the eyes (for sensing
light), the ears (for sensing sound), and so on. These organs absorb the fac-
tual data (such as sights, light, sounds, pressure, and texture) and transfer
it to the brain for processing.
ai technology has the capability and the relevant devices to absorb
any factual data that can be collected by the five human senses. Cameras
absorb light and transfer the data to processors,40 microphones pick up
sounds,41 transducers sense pressures, thermometers sense temperature,
hygrometers sense humidity, and so on. Indeed, advanced technologies are
capable of much more accurate sensing than their corresponding human
organs. For example, cameras may absorb light waves at frequencies that
the human eye cannot detect, and microphones can sense sound waves
that are inaudible to human ears.
How many of us can guess the exact outside temperature and humidity
within 0.1 degrees? What about 0.0001 degrees? And what about relative
humidity? Most of us cannot. By contrast, even the simplest electronic
sensors can perform accurate measurements of these data and transfer the
information to the appropriate processors. ai technology, therefore, is ca-
pable of meeting the requirements of the first stage of awareness, indeed
much better than humans.
The second stage of the awareness process consists of creating a rele
vant general image about these data in the brain, that is, full perception.
Naturally, most ai technologies, robots, and computers do not possess bio-
logical brains, but they possess artificial ones. Most of these “brains” are
embodied in the hardware (processors, memory, and so on) of the technol-
ogy they use. Are these brains capable of creating a relevant general image
out of the factual data they absorb? Humans create the relevant general
image by analyzing the factual data, which enables them to use the infor-
mation, transfer it, integrate it with other information, and act based on
it — in other words, to understand it.42
Consider the example of security robots based on ai technology.43 The
task of these robots is to identify intruders and either alert human troops
(police, army) or stop the intruders on their own. Their sensors (cameras,
microphones) absorb the factual information and transfer it to the pro-
cessors, which are then supposed to identify the intruder. To do so, the
processors analyze the factual data, making sure not to mistake police and
other security personnel in the area for intruders. The analysis is based on
identifying and discriminating between changing sights and sounds. The
robots may compare the shape and color of clothes and use whatever other
attributes they need to make the identification.
Following this short process, the robot assesses the probabilities. If
the probabilities do not produce an accurate identification, the robot ini-
tiates a process of vocal identification and asks the suspicious figure to
identify itself, provide a password, or perform any other action relevant to
the situation. If the intruder answers, the sound is compared with sound
patterns stored in memory. After all the necessary factual data has been
collected, the robot has sufficient information to make a decision, on the
basis of which it can act. Indeed, this robot has created a relevant general
image out of the factual data absorbed by its sensors, which it was able to
use, transfer, and integrate with other information, and then act based on
it — in other words, to understand it.
Consider how a human guard would act in this situation: very likely,
in the same way. The human guard would identify a figure or a sound
as suspicious, whereas the robot would examine a change in the current
image or sound. We prefer robot guards because they work more thor-
oughly than their human counterparts and do not fall asleep on the job.
Both the human guard and the robot use their memories to try to identify
the figure or the sound by comparing it with patterns stored in memory.
But whereas the robot cannot forget figures, humans can. The human
guard is not always sure, but the robot assesses all the probabilities. The
human guard calls out to ask the figure to identify itself, and so does the
robot. Both the human and the robot compare the answer with information
stored in memory — yet another function that the robot carries out more
accurately than its human counterpart. Following all the data collection,
both human and robot guards make a decision. They both understand the
situation.
We say that the human guard was aware of the relevant factual data.
Can we not say the same about the robot guard? There is no reason why
not. Its internal processes are more or less the same as those of the human
guard, except that the robot is more accurate, faster, and works more thor-
oughly than the human guard. When they become aware of a suspicious
figure or sound, they both absorb the data and act accordingly.
It is possible to argue that the human guard may have absorbed much
more factual information than the robot — for example, smells — and
that a human is capable of filtering out irrelevant information, such as
cricket sounds. This argument does not invalidate the analysis just pre-
sented. Robots are also capable of absorbing additional factual data, such
as smells. But as already noted, humans use the attention process to focus
on a portion of the factual data, as Archie did in the example cited earlier.
Although humans are capable of absorbing a wide range of factual data,
this may prove to be a distraction for attending to the task at hand. Robots
may be programmed to ignore such data.
If smells are considered irrelevant for a task of guarding, then robots
will not consider this data. If they were human, we would say that the
robots will not pay attention to this data. Humans filter out irrelevant data
using the process of attention that runs in the background of the human
mind. Robots, however, examine all factual data and eliminate the irrele
vant options only after analyzing the data thoroughly. We may now ask
ourselves which type of guard we humans would prefer: one who uncon-
sciously does not pay attention to factual data, or one who examines all
factual data thoroughly?
In sum, ai technology is also capable of fulfilling the requirements for
the second stage of awareness. And given that the two stages already de-
scribed are the only stages of the awareness process recognized by crimi-
nal law, we may conclude that ai technology is capable of meeting the
awareness requirement in criminal law.
Nevertheless, some of us may have the feeling that something is still
missing before we conclude that machines are capable of awareness. This
may be correct if we consider awareness in its broader sense, as it is used
in psychology, philosophy, or the cognitive sciences. But in this book
we are examining the criminal liability of ai entities, and not the wide
meaning of cognition in psychology, philosophy, the cognitive sciences,
and so on. Therefore, the only standards of awareness that are relevant
here are the standards of criminal law. The other standards are irrele-
vant for the assessment of criminal liability imposed on either humans or
ai entities.
The criminal law definition of awareness is indeed much narrower
than corresponding definitions in the other spheres of knowledge. This is
true for the imposition of criminal liability not only on ai entities, but also
on humans. The definitions of awareness in psychology, philosophy, or
the cognitive sciences may be relevant for the endless quest for machina
sapiens, but not for machina sapiens criminalis, which is based on the
definitions of criminal law.44 This distinction is one of the best illustra-
tions of the difference between machina sapiens and machina sapiens
criminalis.
As noted earlier, awareness itself is very difficult to prove in court, es-
pecially in criminal cases where it must be proved beyond any reasonable
doubt.45 Therefore, criminal law developed two evidential substitutes for
awareness: the presumptions of willful blindness and awareness. Accord-
ing to the willful blindness presumption, the offender is presumed to be
aware of the conduct and circumstances if he or she suspected that they
existed but did not verify that suspicion.46 The rationale of this presump-
tion is that because the “blindness” of the offender to the facts is willful,
he or she is considered to be aware of these facts even if not actually aware
of them.
If the offender wished to avoid the commission of the offense, he or
she would have checked the facts. For example, the rapist who suspects
that the woman does not consent to having sexual intercourse with him
predicts that if he were to ask her, she would refuse, in which case there
would be no doubt about her lack of consent. Therefore, he ignores his
suspicion, and at his interrogation says that he thought she consented be-
cause he heard no objection. The willful blindness presumption equates
unchecked suspicion with full awareness of the relevant conduct or cir-
cumstances. The question is whether this presumption is relevant to ai
technology.
According to the awareness presumption, a person is presumed to be
aware of the possibility of the occurrence of the natural results of his or her
conduct.47 The rationale of this presumption is that humans have the basic
skills necessary to asses the natural consequences of their conduct. For
example, when firing a bullet at someone’s head, the shooter is presumed
to be able to asses the possibility of death as a natural consequence of the
shooting. Humans who lack such skills, permanently or temporarily, have
an opportunity to refute the presumption. The question is whether this
presumption is relevant to ai technology.
Awareness is indeed difficult to prove for humans. But because within
a criminal law context, the processes that comprise awareness can be mon-
itored accurately by ai technology, there is no need for any substitute. In
human terms, this would be the same as if the brain of every individual
were constantly being scanned and every thought recorded. If awareness

of factual data could be identified on such a brain scan, proving awareness
beyond any reasonable doubt would be a simple task.
Given that every act of an ai entity can be monitored and recorded, in-
cluding all processes that comprise awareness within a context of criminal
law, it is entirely achievable to prove the awareness of an ai entity regard-
ing particular items of factual data. Therefore, it is not necessary to prove
the awareness of an ai entity by means of substitutes, because awareness
can be proven directly. But if it were necessary to prove awareness using
substitutes, this could also be accomplished. Strong ai technologies use
algorithms to assess probabilities in order to make reasonable decisions.
A minimum probability defines the feasibility, probability, or “reasonable-
ness” of every event. This praxis can be applied beneficially to awareness
substitutes.
For humans, only a realistic suspicion is considered adequate for will-
ful blindness. Thus, if the probability of a particular factual event occur-
ring is above a certain rate and the ai entity ignores the possibility, this
may be considered willful blindness. The reason a human would not check
such a suspicion is immaterial for the applicability of the willful blindness
presumption, and the same is true for ai entities. So if the argument is that
the ai entity ignores the possibility out of some evil or concealed desire,
as humans may do, the reason for this ignorance is immaterial, just as it is
in the case of humans.
As noted earlier, evil is not one of the components of criminal liabil-
ity. Thus, the reason for ignoring a suspicion (that is, an alternative with
a probability higher than a certain threshold) is immaterial, regardless of
whether the motive is evil. The natural consequence of certain conduct
has a reasonable probability of occurring. A machine can assess reason-
ability quantitatively by weighing the relevant circumstances. Strong ai
technology is capable of identifying the options that humans would call
“natural.” Thus, although proving the awareness of an ai entity is not sub-
ject to the difficulties of proving human awareness, and therefore there is
no need to use the awareness substitutes, these substitutes can be used to
prove the awareness of ai entities.
2.2.3. F ulfillment of the Volitive A spect

The volitive aspect of mens rea contains three levels of will: intent, indif-
ference, and rashness. As already noted, indifference and rashness are
commonly referred to as recklessness. The question is whether ai technol-
ogy is capable of consolidating these levels of will. Because these terms
have different meanings in various scientific spheres (psychology, theol-

ogy, philosophy, law, and so on), to answer this question we must exam-
ine the legal meaning of these terms; specifically, their legal definitions in
criminal law. Even if the legal meaning of these terms differs from other
meanings, given that the question relates to criminal liability, only the
legal meaning is relevant.
Intent is the highest level of will accepted by criminal law. There is
some confusion about the meanings of the terms intent, general intent, and
specific intent.48 General intent is a common term for mens rea. To avoid
confusion with intent, we use the term mens rea and not general intent.
The term intent refers to the highest level of will embodied in the volitive
aspect of mens rea. The term specific intent refers also to the highest level
of will. Intent and specific intent have identical levels of will. But intent
refers to will toward the result components of the factual element, whereas
specific intent refers to motives and purposes, not to results.49 Specific
intent is relatively rare.
Purpose is factual data that are supposed to derive from the conduct.
The mission of the conduct (or its destiny) is to achieve the purpose. Pur-
pose offenses do not require that the purpose actually be achieved, only
that the offender intends to achieve the purpose. For example, “Whoever
says anything for the purpose of intimidating . . .” is an offense that does
not require actual intimidation of anyone, only the presence of such pur-
pose. Motive consists of internal-subjective feeling, and its satisfaction is
the conduct derived from it. For example, in “Whoever performs X out of
hatred . . .” the conduct X satisfies the hatred.
Unless purpose and motive are explicitly required by the offense, they
are immaterial for the imposition of criminal liability.50 For example, (1)
A killed B out of hatred, and (2) A killed B out of mercy. These are both
considered murders because the offense of murder does not require spe-
cific intent to achieve a set purpose or any motives. But if the offense had
explicitly required specific intent toward a set purpose or any motives,
proving the specific intent would have been a condition for the imposi-
tion of criminal liability in that offense. Again, the requirement of specific
intent is relatively rare.
Given that the only difference between intent and specific intent in-
volves their objects (results versus motives and purposes), and that the
level of will of both is identical, the following analysis of intent applies to
specific intent as well. Intent is defined as “aware will” that accompanies
the commission of conduct, the results of which will occur. In the defini-
tion of specific intent, results are replaced by motives and purposes as fol-
lows: aware will that accompanies the commission of conduct, the motive
for which is satisfied or the purpose of which is achieved.
Intent is an expression of positive will, that is, the will to make a fac-
tual event occur.51 Although there are higher levels of will than intent (for
example, lust, longing, or desire), in criminal law intent has been accepted
as the highest level of will required for the imposition of criminal liabil-
ity.52 Consequently, no offense requires a higher level of will than intent,
and if intent is proven, it satisfies all other levels of will, including reck-
lessness (indifference and rashness).
We must distinguish between aware will and unaware will. Unaware
will is an internal urge, impulse, or instinct of which the person is not
aware and which is naturally uncontrollable. Individuals cannot control
their will unless they are aware of it. Being aware of one’s will does not
guarantee the ability of controlling it; but controlling the will requires
awareness of it and the activation of conscious processes in the mind to
cause the relevant activity to start, cease, or continue without interference.
Imposing criminal liability based on intent requires an individual who is
aware of his will so that he may control it.
Intent does not require some abstract aware will. The aware will must
be focused on certain targets: results, motives, or purposes. Intent is an
aware will focused on these targets. For example, the murderer’s intent is
an aware will focused on causing the victim’s death. For the intent to be
relevant to the imposition of criminal liability, this will must exist simul-
taneously with the commission of the conduct and must accompany it. If
a person kills another by mistake, and then, after the victim’s death, devel-
ops the will to kill the victim, this is not intent. Only when the murder is
committed simultaneously with the existence of the relevant will can the
will be considered intent.
Proving intent is much more difficult than proving awareness. Al-
though both are internal processes of the human mind, intent relates to
future factual situations, and awareness relates to current facts. Aware-
ness is rational and realistic, but intent is not necessarily so. For example,
a person may have the intention of becoming a fish, but she cannot have
an awareness of being a fish. Because of the serious difficulties in proving
intent, criminal law developed evidentiary substitutes for it. The most
common substitute is the foreseeability rule (dolus indirectus), a legal pre-
sumption designed to prove the existence of intent.
According to the foreseeability rule presumption, the offender is pre-
sumed to intend the occurrence of the results if during commission of the
conduct, with full awareness, he has foreseen the occurrence of the results
as highly probable.53 If the object of the results is replaced with purpose,

this presumption applies also to specific intent.54 The rationale for this
presumption is as follows: believing that the probability of a specific fac-
tual event following from the conduct is extremely high, committing the
conduct expresses the offender’s desire that the factual event should occur.
For example, A holds a loaded gun pointed at B’s head. A knows that
if she pulls the trigger, there is a high probability that the factual event of
B’s death will occur. A pulls the trigger. A then argues in court that she
did not want B’s death to occur, and therefore the required component
of intent has not been fulfilled. If the court exercises the foreseeability
rule presumption, A is presumed to have intended the occurrence of the
result because in A’s assessment, the result of death was highly probable
and she acted accordingly, presumably because she wanted these results
to occur.55 This presumption is commonly used in courts to prove intent.
Indeed, unless the defendant confesses explicitly during the interrogation
to having had the intent, the prosecution prefers to prove the intent using
this presumption.
Is an ai robot capable of having intent in the context of criminal law?56
Because will is a vague and general term, even in criminal law, an ai en-
tity’s capability of having intent must be examined through the lens of the
foreseeability rule presumption. Indeed, this is the core reason for using
the presumption in proving human intent. Two conditions must be met
by this rule: (1) occurrence of the results must be foreseeable with a very
high probability, and (2) the conduct must have been committed with full
awareness.
As already noted, strong ai entities are capable of assessing probabili-
ties of the occurrence of factual events and acting accordingly. For exam-
ple, chess-playing computers can analyze the current status of the game
based of the location of the pieces on board. These computers review all
possible options for the next move, and for each option they review all the
possible responses of the other player. Then, for each response they again
review all possible responses, and so on until the final possible move that
ends the game with a victory for one of the players. Each possible move is
assessed for its probability, based on which, the computer decides on its
next move.57
If the player were human, we would say that he has an intention of
winning the game. We would not know with certainty that he has such
an intention, but the course of his conduct is consistent with the fore-
seeability rule presumption. An ai computer programmed to play chess
has a goal-driven behavior of winning chess games, as do human chess
players. If we can say that human chess players have the intention of win-
ning chess games, it seems that we can say the same thing about ai chess.
Analysis of their course of conduct in the relevant situations is entirely
consistent with the foreseeability rule presumption.
Any entity, human or ai, that examines several options for conduct
and then makes an aware decision to commit one of them, while assess-
ing the probability that a given factual event would result from this con-
duct with a high degree of probability, is considered to be foreseeing the
occurrence of that factual event. In the case of a chess game, this seems
quite normal to us. But there is no substantial difference, in this context,
between playing chess for the purpose of winning and committing any
other conduct for the purpose of bringing about some result. If the result
and the conduct together form a criminal offense, we enter the territory
of criminal law.
When a computer assesses the probability that a factual event (win-
ning a chess game, for example, or killing or injuring a person) will result
from its conduct and chooses accordingly to commit the relevant conduct
(moving chess pieces on the board, pulling a trigger, or moving its hydrau-
lic arm in the direction of a human body), the computer meets the required
conditions for the foreseeability rule presumption. The computer is there-
fore presumed to have the intention that the results should occur. This is
exactly how the court examines the offender’s intent in most cases, when
the offender does not confess.
One may ask, what is considered to be a very high probability in this
context? The answer is the same as for humans, the only difference being
that the computer can assess the probability more accurately than a human
can. For example, one points a loaded gun at someone’s head. A human
shooter evaluates that the probability of the victim’s death is high as a
result of pulling the trigger, but most humans are incapable of assessing ex-
actly how high the probability is. If the entity is a computer, it can assesses
the exact probability of occurrence based on the factual data to which it
is exposed to (the distance of the victim from the gun, the direction and
velocity of the wind, the mechanical condition of the gun, and so on).
If probability is assessed to be high, the computer is required to act
accordingly, and in full awareness to commit a conduct that advances the
relevant factual event. As already noted, ai technology has the capability
to consolidate awareness to factual data.58 Commission of the conduct is
considered as factual data, therefore ai technology can meet both condi-
tions needed to prove the foreseeability rule presumption. This presump-
tion is considered to be an absolute legal presumption (praesumptio juris
et de jure); therefore, if its conditions are met, there is no way of refuting

its conclusion (that is, having intent).
This analysis of the foreseeability of ai technology relies on strong ai ,
the more specific requirement being the advanced capability of assessing
probabilities for the purpose of decision making. Because strong ai tech-
nologies have this capability, they are therefore capable of intent within
the context of criminal law. Thus, ai technology can meet the intent re-
quirement in criminal law.
Nevertheless, some of us may still have the feeling that something is
missing before we conclude that machines are capable of intent. This may
be correct if we consider intent in its broader sense, as used in psychol-
ogy, philosophy, or the cognitive sciences. But here we are examining the
criminal liability of ai entities, and not the wider meaning of cognition.
Therefore, the only standards of intent that are relevant to our examination
are the standards of criminal law. Other standards are irrelevant for the
assessment of criminal liability imposed on either humans or ai entities.
The criminal law definition of intent (and foreseeability) is indeed
much narrower than corresponding definitions in other spheres of knowl-
edge. This is true for the imposition of criminal liability not only on ai
entities, but also on humans. The definitions of intent in psychology, phi-
losophy, or the cognitive sciences may be relevant for the endless quest
for machina sapiens, but not for machina sapiens criminalis, which is
based on the definitions of criminal law.59 This distinction is one of the
best illustrations of the difference between machina sapiens and machina
sapiens criminalis.
The evidence concerning the intent of the ai entity is based on the
ability to monitor and record all the activities of the computer. The ac-
tivities of the computer with regard to each stage in the consolidation of
intent or foreseeability are monitored and recorded. Assessing probabili-
ties and making relevant decisions based on the assessment are part of the
computer’s activities, so there is always direct evidence for proving the
criminal intent of an ai entity if the proof is based on the foreseeability
rule presumption.
If intent is proven, either directly or through the foreseeability rule
presumption, then all other forms of volition may be proven accordingly.
Because recklessness, consisting of indifference or rashness, represents
a lower degree of will, it may be proven either by proving recklessness
directly or by proving intent. There are no offenses that require only reck-
lessness and nothing else. Because criminal liability requirements func-
tion as a matrix of minimal requirements, as noted earlier,60 every require-
ment represents the minimum condition for the imposition of criminal

liability. The prosecution may choose between proving the given require-
ment or any higher (but not lower) one.
Consequently, offenses that require recklessness as their mental ele-
ment requirement can be satisfied by proving either intent (directly or
through the foreseeability rule presumption) or recklessness. Thus, if the
given ai technology has the foreseeability capabilities already described, it
can meet the mental element requirements of both intent and recklessness
offenses. But ai technology can also meet the recklessness requirement
directly. Analogously, if the capability for intent is present, the capability
for recklessness, which is a lower one, is also present.
Indifference is a higher level than recklessness, and it represents a
fully aware state of volitional neutrality toward the occurrence of the fac-
tual event. For the indifferent person, the significance of the factual event
occurring and not occurring is exactly the same. This has nothing to do
with the actual probability of the occurrence of the factual event, only with
the offender’s internal volition toward such occurrence. For example, A
and B are playing Russian roulette.61 When it is B’s turn, A is indifferent to
the possibility of B’s death; A does not care whether B lives or dies.
For indifference to be considered as such in the context of criminal
law, it must be aware; that is, the offender must be aware of the various
options and have no preference for any of them.62 The decision-making
process in strong ai technology is based on assessing probabilities, as we
have discussed. When the computer makes a decision to act in a certain
way, but this decision does not take into consideration the probability that
a given factual event can occur, the computer is indifferent to the occur-
rence of that factual event.
In general, complicated decision- making processes in ai technol-
ogy are characterized by many factors to be considered. In similar situ-
ations, humans tend to ignore a portion of the factors and not consider
them. Computers do the same. Some are programmed to ignore certain
factors, but strong ai technology is capable of learning to ignore factors —
otherwise, the decision-making process would be impossible. The learn-
ing process used by strong ai technology, machine learning, is inductive
learning from examples. The more examples analyzed, the more effective
the learning.63 Humans call this “experience.”
Whenever the decision maker is an ai entity, the decision-making pro-
cess is monitored, so there is no evidential problem in proving the indif-
ference of the ai entity toward the occurrence of a given factual event. The
awareness of the ai entity to the possibility of the occurrence of the event
is monitored, together with the factors that were taken into consideration
in the decision-making process. This data provides direct proof of the in-
difference of the ai entity. Indifference can also be proven through the
foreseeability rule presumption, as noted earlier.
Rashness is the lower level of recklessness. It consists of aware volition
for the relevant factual event not to occur, but at the same time the com-
mitment of aware and unreasonable conduct that causes it to occur. Al-
though the occurrence of the factual event is undesired by the rash person,
she conducts herself with unreasonable risk so that the event eventually
occurs. Rashness is considered a degree of will, because if the rash of-
fender had not wanted the event to occur at any price, she would not have
taken unreasonable risks in her conduct. This is the major reason that
rashness is part of the volitive aspect of mens rea. Otherwise, negative will
alone does not justify criminal liability.
For example, someone drives on a narrow road behind a slow truck.
It is a two-way road divided by a solid white line. The driver is in a hurry
and eventually decides to pass the truck, crossing the solid line. She does
not intend to kill anyone; all she wants is to pass the truck. But a motor
cyclist happens to come from the opposite direction. The driver is unable
to avoid the motorcyclist, and kills him. If the driver had wanted the mo-
torcyclist to be dead, she would have been criminally liable for murder,
but this was not the case. Nor would it be true to say that she was indiffer-
ent to the death of the motorcyclist. She was rash, however. She did not
want to hit the motorcycle, but she took unreasonable risks that resulted
in death. She is therefore criminally liable for manslaughter.
In the context of criminal law, rashness requires awareness. The of-
fender is required to be aware of the possible options, to prefer not to cause
the occurrence of the resulting event, but to commit conduct that is unrea-
sonable if she were to avoid the occurrence of the event.64 The decision-
making process of strong ai technology is based on assessing probabilities,
as we have discussed. When the computer makes a decision to act in a
certain way, but it does not weigh one of the relevant factors as sufficiently
significant, it is considered to be rash regarding the occurrence of a given
factual event. In our example, if the car driver were replaced by a driving
computer, and if the computer calculated the probability of colliding with
the motorcycle by crossing the solid line to be low, the decision to pass
would be considered rash.
As already noted, complicated decision-making processes in ai are
characterized by many factors to be considered. In situations like this,
humans sometimes miscalculate the weight of some factors. So do com-
puters. Humans reinforce their decisions by hopes and beliefs, but com-
puters do not use such methods. Some computers are programmed to
weigh certain factors in a certain way, but strong ai technology can learn
to weigh factors correctly and accurately. Again, this learning by strong
ai is called machine learning.
Whenever the decision maker is an ai entity, the decision-making pro-
cess is monitored; therefore, there is no shortage of evidence in proving
rashness on the part of an ai entity regarding the occurrence of a given fac-
tual event. The ai entity’s awareness of the possibility of the occurrence of
the event is monitored, together with the factors that were taken into con-
sideration in the decision-making process, and also the actual weight they
were assigned in the given decision. This data provides direct proof of
the rashness of the ai entity. As noted earlier, rashness can also be proven
through the foreseeability rule presumption.
Therefore, all components of the volitive aspect of mens rea apply to
ai technology and can be proven in court. So the question is, who is to be
held criminally liable for the commission of this type of offense?
2.3. C ri minall y L iable Entities for Intention al ai Offe nses

In general, imposition of criminal liability for mens rea offenses (“in-
tentional offenses”) requires the fulfillment of both factual and mental
elements of these offenses. Humans are involved in the creation of ai
technology and entities, in their design, programming, and operation. Con-
sequently, when an ai entity meets the factual and mental elements of the
offense, who is to be held criminally liable? The possible answers are the
ai entity, the humans involved with the ai entity, or both. Let us examine
each of these answers.
2.3.1. ai E ntity L iability

If a human offender meets both the factual and mental element require-
ments of an offense, criminal liability for that offense is imposed. In
doing so, the court does not need to investigate whether the offender was
“evil,” or whether some other attribute characterized the commission of
the offense. The fulfillment of these requirements is the only condition
for imposing criminal liability. Other information may affect the punish-
ment, but not the criminal liability. We prefer our criminal legal system
to act in this way because evil and good are matters of perspective.65 Dif-
ferent people with differing perspectives regard these terms in different
ways.66
The factual and mental elements are neutral in this context, and do
not necessarily have evil or good content.67 Meeting their requirements is
much more “technical” than the detection of evil would be. For example,
we prohibit murder. Murder is causing the death of a human with the
awareness and intent of causing death. If an individual factually causes
another person’s death, the factual element requirement is fulfilled. If the
conduct was committed with full awareness and intent, the mental ele-
ment is also fulfilled, and the individual is criminally liable for murder,
unless one of general defenses is applicable (for example, self-defense or
insanity). The reason for the murder is immaterial for the imposition of
criminal liability, and it makes no difference whether the murder was com-
mitted out of mercy (euthanasia) or out of villainy.
This is the way that criminal liability is imposed on human offenders.
If we embrace this standard in relation to ai entities, we can impose crimi-
nal liability on them as well. This is the basic idea behind the criminal
liability of ai entities, and it differs from their moral accountability, social
responsibility, or even civil legal personhood.68 The narrow definition of
criminal liability makes it possible for the ai entity to become subject to
criminal law. Nevertheless, the reader may feel that something is still miss-
ing in this analysis, that perhaps it falls short. We try to refute these feel-
ings using rational arguments.
One such feeling may be that the capacity of an ai entity to follow a
program is not sufficient to enable the system to make moral judgments
and exercise discretion, although the program may contain a highly elabo-
rate and complex system of rules.69 This feeling may relate eventually to
the moral choice of the offender. The deeper argument is that no formal
system can adequately make the moral choices that may confront an of-
fender. Two answers apply to this argument.
First, it is not certain that formal systems are morally blind. There are
many types of morality and moral values.70 Teleological morality, such as
utilitarianism, deals with the utility values of conduct. These values may
be measured, compared, and decided upon based on quantitative compari-
sons. For example, an ai robot controls a heavy wagon that malfunctions,
leaving two possible paths for the wagon to follow. The robot calculates
the probabilities, and determines that following one path will result in the
death of one person, and following the other path will result in the death
of fifty people. Teleological morality would direct any human to choose
the first path, and the person who followed that path would be considered
a moral person. The robot can be directed to act in accordance with this
morality. The robot’s morality is dictated by its program, which makes it
evaluate the consequences of each possible path. Do humans not act the
same way?
Second, even if the first answer is not convincing, and formal sys-
tems are considered incapable of morality of any kind, criminal liability
is still neither dependent on any morality nor fed by it. Morality is not
even a condition for the imposition of criminal liability. Criminal courts
do not asses the morality of human offenders when imposing criminal li-
ability. The offender may be highly moral in the court’s view, and still be
convicted, as in the case of euthanasia, for example. Alternatively, the of-
fender may be very immoral in the eyes of the court, and still be acquitted,
as in the case of adultery. Given that no morality of any type is required
for the imposition of criminal liability on human offenders, it should not
be a consideration when ai entities are involved, either.
Some may feel that ai entities are not humans, and that criminal li-
ability is intended for humans only because it involves constitutional
human rights that only humans may have.71 In this context, it is immate-
rial whether the constitutional rights refer to substantial rights or to pro-
cedural rights. The answer to this argument is that perhaps criminal law
was originally designed for humans, but since the seventeenth century, it
has not been applied to humans exclusively, as we have discussed.72 Cor-
porations, which are nonhuman entities, are also subject to criminal law,
and not only to criminal law. For the past four centuries, criminal liability
and punishments have been imposed on corporations, albeit with some
occasional adjustments.
Others may argue that although corporations have been recognized as
subject to criminal law, the personhood of ai entities should not be recog-
nized because humans have no interest in recognizing it.73 This argument
cannot be considered applicable in an analytical legal discussion on the
criminal liability of ai entities. There are many cases in everyday life in
which the imposition of criminal liability does not benefit human soci-
ety, and still criminal liability is imposed. A famous example comes from
Immanuel Kant, who is interpreted as believing that if the last person on
Earth is an offender, he should be punished even if it leads to the extinc-
tion of the human race.74 Human benefit has not been recognized as a valid
component of criminal liability.
Finally, some may feel that the concept of awareness presented here
is too shallow for ai entities to be called into account, blamed, or faulted
for the factual harm they may cause.75 This feeling is based on confusion:
the concept of awareness and consciousness in psychology, philosophy,
theology, and cognitive sciences is confused with the concept of awareness
in criminal law. In other spheres of knowledge, we lack a clear notion of

what awareness is, making it impossible to offer coherent answers to the
question of ai capability of awareness. Whenever such answers are given,
they are usually based on intuition rather than on science.76
Criminal law, however, must be accurate. Based on definitions pro-
vided by criminal law, people may be jailed for life, and they may lose
their property or even their lives. Therefore, criminal law definitions must
be proven beyond any reasonable doubt. Imposition of criminal liabil-
ity for mens rea offenses is based on awareness as the major and domi-
nant component of the mental element requirement. The fact that the term
awareness in psychology, philosophy, theology, and cognitive sciences has
not been developed adequately in order to create an accurate definition
does not exempt criminal law from developing such a definition of its own
for the purpose of imposing criminal liability.
The criminal law definition of awareness, like all other legal defini-
tions, may differ significantly from its common meaning or from its mean-
ing in psychology, philosophy, theology, and cognitive sciences. Criminal
law definitions are designed and adapted to meet the needs of criminal law
and nothing more. These definitions represent the requirements necessary
for the imposition of criminal liability. They also represent the minimal
conditions, both structural and substantive, based on the concept of crimi-
nal law as a matrix of minimal requirements, as discussed earlier.77
Consequently, the definitions of criminal law, including the definition
of awareness, are relative and relevant only to criminal law. Because they
are formed in conformity with the matrix of minimal requirements, they
may be regarded as shallow. But this is true only when examined from the
perspectives of psychology, philosophy, theology, and cognitive sciences,
not from that of criminal law. If a significant development were to occur
in those scientific spheres with regard to awareness, the criminal law may
embrace newer and more comprehensive definitions of awareness in-
spired by these scientific discoveries. For the time being, there are none.
Criminal law definitions, including that of awareness, were originally
designed for humans. For centuries people were indicted, convicted, and
acquitted based on these definitions. Since the seventeenth century, these
definitions have been adopted and adapted to the incrimination of non
human entities as well, namely corporations, which have also been in-
dicted, convicted, and acquitted based on these definitions. Whenever
the criminal law definitions were changed, they were changed for both
humans and corporations in the same way. In the twenty-first century, the
same definitions are required for the incrimination of ai entities.
Examination of these definitions reveals that they are applicable to

ai technology, which can meet the requirements of criminal law without
a single change to these definitions. Does that suddenly make them “shal-
low,” when the same definitions have been used by criminal law systems
worldwide for centuries and were considered adequate and almost un
assailable? If they are good enough for humans and corporations, why
would not they be good enough for ai technology?
If criticism claimed that criminal law definitions are too shallow and
must therefore be radically changed for humans, corporations, and ai tech-
nology, the argument might be acceptable. But when these definitions are
called shallow only in relation to ai technology, the argument cannot be
considered serious or applicable. Such criticism originates in the endless
quest for machina sapiens, and it is part of that quest. But because crimi-
nal law is a limited field, the quest for machina sapiens criminalis ended
successfully when ai entities showed that they are capable of complying
with the definitions of criminal law.
This answer applies not only to the criticism regarding the awareness
of ai entities in criminal law, but also to the intentionality of ai entities.78
We can now summarize that the criticism of the idea of ai entities’ crimi-
nal liability generally relates to two points: (1) an absence of attributes that
are not required for the imposition of criminal liability (for example, soul,
evil, and good); and (2) the shallowness of criminal law definitions from
the perspective of science other than criminal law (for example, psychol-
ogy, philosophy, theology, and cognitive sciences). Both points of criticism
can be answered methodically.
The first criticism can be refuted by pointing out the structural aspect
of the matrix of criminal law, and noting that none of the absent attributes
is required for the imposition of criminal liability on humans, corpora-
tions, and ai entities alike. The second criticism is answered through the
substantive aspect of the matrix of criminal law, where definitions of legal
terms from outside the sphere of criminal law are entirely irrelevant for
settling the question of criminal liability. This opens the door for the im-
position of criminal liability on ai entities as direct offenders.
Acceptance of the idea of the criminal liability of ai entities for mens
rea offenses does not end in the commission of a certain offense by ai
entities as principal offenders. The commission of mens rea offenses can
also take place through complicity. No less than mens rea is required for
the imposition of criminal liability on accomplices, because there is no
complicity through negligence or strict liability.79 A joint perpetrator may
be considered as such only in relation to mens rea offenses. Other general
forms of complicity (for example, incitement and accessoryship) also re-

quire at least mens rea.
In sum, because all general forms of complicity require at least mens
rea, an ai system may be considered as an accomplice only if it actually
formed mens rea. Making it possible to impose criminal liability on ai en-
tities as direct offenders also recognizes the possibility of ai entities being
accomplices, joint perpetrators, inciters, accessories, and so on, as long as
both factual and mental element requirements are met in full. But these are
not the only possible criminal liabilities that can be imposed on ai enti-
ties. Two additional ways are possible: perpetration-through-another and
probable consequence liability.
2.3.2. H uman L iability: P erpetration-through-A nother

Humans, corporations, and ai technology alike may be used as instru-
ments in the commission of offenses, regardless of their legal personhood.
For example, one person threatens another that he will be killed if he does
not commit a certain offense. Having no choice, the threatened person
commits the offense. Who is considered criminally liable for the com-
mission of that offense? The threatening person? The threatened person?
Both? In the context of the criminal liability of ai entities, this question
arises when the ai entity is used as an instrument by another entity. In the
future, perhaps, the other way around may be relevant as well . . .
For situations of this type, criminal law has created the general form
of criminal liability called perpetration-through-another. This may be de-
fined as aware execution of a criminal plan through instrumental use of
another person, who participates in the commission of the offense as an
innocent agent or semi-innocent agent. Perpetration-through-another is a
late development of vicarious liability in a law of complicity. Vicarious
liability has been recognized in both criminal and civil law since ancient
times, and it is based on the ancient concept of slavery.80
The master, who was a legal entity and possessed legal personhood,
was liable not only for his own conduct but also for that of all his subjects
(slaves, workers, family, and so on). When one of these subjects commit-
ted an offense, it was considered as if the master himself had committed
the offense, and the master was obligated to respond to the indictment
(respondeat superior). The legal meaning of this obligation was that the
master was criminally liable for offenses physically committed by his sub-
jects. The rationale for this concept was that the master should enforce
the criminal law among his subjects. If the master failed to do so, he was
personally liable for the offenses committed by his subjects.
Because the master’s subjects were considered his property, he was

liable for the harms committed by them, under both criminal and civil law.
A subject was considered an organ of the master — his long arm. The legal
maxim that governed vicarious liability stated that whoever acts-through-
another is considered to be acting for himself (qui facit per alium facit per
se). The physical appearance of the commission of the offense was imma-
terial for the imposition of criminal liability in this context.
This legal concept was accepted in most ancient legal systems. Based
on it, Roman law developed the function of the father of the family (pater-
familias), who was responsible for any crime or tort committed by mem-
bers of the family, its servants, guards, and slaves.81 Consequently, the
father of the family was also responsible for the prevention of criminal
offenses and civil torts among his subjects. His incentive for doing so was
his fear of criminal or tort liability for actions committed by members of
his household. The legal concept of vicarious liability was absorbed into
medieval European law.
The concept of vicarious liability was formally and explicitly accepted
in English common law in the fourteenth century,82 based on legislation
enacted in the thirteenth century.83 Between the fourteenth and seven-
teenth centuries, English common law amended the concept and ruled
that the master was liable for the servants’ offenses (under criminal law)
and torts (under civil law) only if he explicitly ordered the servants to
commit the offenses, explicitly empowered them to do so, or consented to
their doing so before the commission of the offense (ex ante), or after the
commission of the tort (ex post).84
After the end of the seventeenth century, this firm requirement was
replaced by a much weaker one. Criminal and civil liability could be im-
posed on the master for offenses and torts committed by the servants even
if the orders of the master were implicit or the empowerment of the ser-
vants was general.85 This was the result of an attempt by English common
law to deal with the many tort cases against workers at the dawn of the first
industrial revolution in England and of the commercial developments of
that time. The actions committed by the master’s workers were considered
to be the actions of the master because he enjoyed their benefits. And if the
master enjoyed the benefits of these actions, he should be legally liable for
the harm that might be caused by them, under both criminal and civil law.
In the nineteenth century, the requirements were further weakened,
and it was ruled that if the worker’s actions were committed through or as
part of the general course of business, the master was liable for them even
if no explicit or implicit orders had been given. Consequently, the defense
argument of the worker having exceeded his authority (ultra vires) was
rejected. Thus, even if the worker acted in contradiction to the specific
order of a superior, the superior was still liable for the worker’s actions if
they were carried out in the general course of business. This approach was
developed in tort law, but the English courts did not restrict it to tort law,
and applied it to criminal law as well.86
Thus, vicarious liability was developed under specific social condi-
tions, under which only individuals of the upper classes had the required
competence to be considered legal entities. In Roman law, only the father
of the family could become a prosecutor, plaintiff, or defendant. When the
concept of social classes began to fade, in the nineteenth century, vicarious
liability faded away with it.
In criminal law at the beginning of the nineteenth century, the cases
of vicarious liability fell into three main types. The first was classic com-
plicity. If the relations between the parties were based on real cooperation,
they were classified as joint perpetration even if they had an employer-
employee or some other hierarchical relation. But if within the hierarchi-
cal relations, information gaps between the parties or the use of power
made one party lose its ability to commit an aware and willed offense, the
act was no longer considered joint perpetration. The party that had lost the
ability to commit an aware and willed offense was considered an innocent
agent, functioning as a mere instrument in the hands of the other party.
The innocent agent was not criminally liable, and the offense was con-
sidered perpetration-through-another, where another party had full crimi-
nal liability for the actions of the innocent agent.87 This was the basis for
the emergence of perpetration-through-another from vicarious liability,
and it became the second type of criminal liability derived from vicarious
liability. The third type accounted for the core of the original vicarious
liability. In most modern legal systems, this type is embodied in specific
offenses and not in the general formation of criminal liability. Since the
emergence of the modern law of complicity, the original vicarious liability
is no longer considered a legitimate form of criminal liability.
Since the end of the nineteenth century and the beginning of the twen-
tieth, the concept of the innocent agent has been widened to include par-
ties with no hierarchical relations between them. Whenever a party acts
without awareness of its actions or without will, it is considered an inno-
cent agent. The acts of an innocent agent could be the results of another
party’s initiative (for example, using the innocent agent through threats,
coercion, misleading, or lies); or the results of another party’s abuse of an
existing factual situation that eliminates the awareness or will of the in-
nocent agent (for example, abuse of a factual mistake, insanity, intoxica-

tion, or infancy).
During the twentieth century, the concept of perpetration-through-
another has been applied also to semi-innocent agents, typically a neg-
ligent party that is not fully aware of the factual situation although any
other reasonable person could have been aware of it under the same cir-
cumstances. Most modern legal systems accept the semi-innocent agent as
part of perpetration-through-another, so that the other party is criminally
liable for the commission of the offense, and the semi-innocent agent is
criminally liable for negligence.
If the legal system contains an appropriate offense of negligence (that
is, the same factual element requirement, but a mental element of negli-
gence instead of awareness, knowledge, or intent), the semi-innocent agent
is criminally liable for that offense. If no such offense exists, no criminal
liability is imposed, although the other party is criminally liable for the
original offense. For the criminal liability of perpetration-through-another,
the factual element may be fulfilled by the innocent agent, but the mental
element requirement must be fulfilled actually and subjectively by the
perpetrator-through-another, including with regard to the instrumental use
of the innocent agent.88
So the question arises, if an ai entity is used by another (human, cor-
poration, or another ai entity) as an instrument for committing an offense,
how would criminal liability be divided between them? Perpetration-
through-another does not consider the ai entity that physically committed
the offense as possessing any human attributes. The ai entity is considered
an innocent agent, but we cannot ignore an ai entity’s capabilities to physi-
cally commit the offense. These capabilities are not sufficient to consider
the ai entity as the perpetrator of the offense, because the entity lacks the
required awareness or will.
The capabilities of the ai entity to physically commit the offense re-
semble the corresponding capabilities of a mentally limited person, such
as an infant,89 a person who is mentally incompetent,90 or one who lacks a
criminal state of mind.91 Legally, when an offense is committed by an in-
nocent agent (an infant,92 a person who is mentally incompetent,93 or one
who lacks a criminal state of mind to commit an offense94), no criminal
liability is imposed on the physical perpetrator. In such cases, that person
is regarded as an instrument, albeit a sophisticated one, while the party
orchestrating the offense (the perpetrator-through-another) is the actual
perpetrator as a principal in the first degree, and is held accountable for
the conduct of the innocent agent.
The perpetrator’s liability is determined on the basis of the conduct of

the “instrument”95 and of his mental state.96 The derivative question re-
garding ai entities is, who is the perpetrator-through-another? The answer
is any person who makes an instrumental use of the ai technology for the
commission of the offense. In most cases, this person is human, and may
be the programmer of the ai software, its operator, or its end user. A pro-
grammer may design the ai entity in order to commit offenses through it.
For example, a programmer designs software for a robot for use in a factory,
and its software is programmed to torch the factory at night when no one
is present. The robot committed the arson, but the programmer is consid-
ered the perpetrator.
The end user did not program the software, but uses the ai entity, in-
cluding its software, for her own benefit, which is expressed by the com-
mission of the offense. For example, a user purchases a household robot
designed to execute any order given by its master. The robot identifies the
user as the master, and the master orders the robot to assault any intruder
to the house. The robot executes the order exactly as instructed. This situ-
ation is not different from that of a person who orders her dog to attack
any trespasser. The robot commits the assault, but the user is considered
the perpetrator.
In both scenarios, the actual offense was physically committed by an
ai entity. Because the programmer and the user did not perform any action
that conforms to the definition of a specific offense, they do not meet the
factual element requirement of the offense. The liability in perpetration-
through-another considers the physical actions committed by the ai entity
as if they had been committed by the programmer, the user, or any other
person instrumentally using the ai technology. The legal basis for this
criminal liability is the instrumental use of the ai entity as an innocent
agent.97 No mental attribute required for the imposition of criminal liabil-
ity is credited to the ai entity.98
When programmers or operators use an ai entity instrumentally, the
commission of an offense by the ai entity is attributed to them. The mental
element required for the offense already exists in their minds. The pro-
grammer had criminal intent when he ordered the commission of the
arson, and the user had criminal intent when she ordered the commission
of the assault, even if these offenses were physically committed by a robot,
which is an ai entity. When an end user makes instrumental use of an in-
nocent agent to commit an offense, the end user is considered to be the
actual perpetrator of the offense.
Perpetration-through-another does not attribute any mental or human
mental capabilities to the ai entity. Therefore, there is no legal difference

between an ai entity, a screwdriver, and an animal that are instrumentally
used by the perpetrator. When a burglar uses a screwdriver to pry open a
window, he does so instrumentally, and the screwdriver is not criminally
liable. The “action” of the screwdriver is in practice the burglar’s. This is
the same legal situation as when using an animal instrumentally. An as-
sault committed by a dog by order of its master is considered as an assault
committed by the master.
This type of criminal liability may be suitable for two kinds of sce-
narios. In the first scenario, an ai entity, even strong ai technology, is used
to commit an offense without taking advantage of the advanced capabili-
ties of the entity. In the second scenario, a weak version of ai technology
is used, which lacks the advanced capabilities of modern ai entities. In
both scenarios the ai entity is used instrumentally, but nevertheless, in
both cases the ai entity is used because of its ability to execute an order
to commit the offense. A screwdriver cannot execute such an order, but a
dog can. A dog cannot execute a complicated order, but an ai entity can.99
Liability of perpetration-through-another is not applicable when an
ai entity makes the decision to commit an offense based on its own ac-
cumulated experience or knowledge, or based on advanced calculations
of probabilities. This liability is not suitable when the software of the ai
entity was not designed to commit a specific offense, which was commit-
ted by the ai entity nevertheless. This type of liability does not apply either
when the ai entity functions not as an innocent, but as a semi-innocent
agent.100 Semi-innocent agents lack the mens rea component but have a
lower mental element component, such as negligence or strict liability, as
we will discuss later.101
Liability of perpetration-through-another may still be applicable, how-
ever, when a programmer or user makes instrumental use of an ai entity,
but without using its advanced capabilities. The legal result of applying
this liability is that the programmer and the user are criminally liable for
the offense committed, whereas the ai entity has no criminal liability of
any type.102 This is not significantly different from treating ai personhood
as mere property, albeit with sophisticated skills and capabilities.103 If the
ai entity is considered a semi-innocent agent, that is, it fulfills the negli-
gence or strict liability requirements, it is criminally liable for offenses of
negligence or strict liability if there are recognized by the criminal law.
2.3.3. Joint H uman and ai E ntity L iability:

P robable C onsequence L iability
The first type of criminal liability, presented earlier, treated the ai entity
as the perpetrator of the offense.104 The second treated the ai entity as a
mere instrument in the hands of the legally-considered perpetrator.105 This
second type of liability is not the only one possible to describe the legal
relations between humans and ai entities in the commission of the offense.
The second type dealt with adhered ai entity. But what if the ai entity,
which was not programmed to commit the offense, calculated its actions,
which then turn out to be considered an offense? The question here con-
cerns the human liability rather than that of the ai entity.
For example, the programmer of a sophisticated ai entity designs it not
to commit certain offenses. In the beginning, after being placed in opera-
tion, the ai system commits no offenses. In time, machine learning through
induction widens the repertoire of the ai entity, which pursues new paths
of activity. At some point, an offense is committed. In another, somewhat
different example, the programmer designs the ai system to commit one
given offense. As expected, the ai system commits the offense, but then it
deviates from the original plan of the programmer and continues its delin-
quent activities. The deviation may be quantitative (more offenses of the
same type), qualitative (more offenses of a different type), or both.
If the programmer had designed the entity from the beginning to commit
the additional offenses, this would have been considered perpetration-
through-another at most. But this is not what the programmer did. If the
ai system consolidates both factual and mental elements of the additional
offenses, it is criminally liable according to the first type of liability. But
the question here concerns the criminal liability of the programmer. This
is the main issue of the third type of liability, which we will discuss.
The most appropriate criminal liability in these cases is the probable
consequence liability. Originally, probable consequence liability in crimi-
nal law was related to the criminal liability of parties to a criminal offense
that has been committed in practice, but that was not part of the original
criminal plan. For example, A and B plan to commit a bank robbery. Ac-
cording to the plan, A’s role is to break into the safe, and B’s role is to
threaten the guard with a loaded gun. During the robbery the guard resists,
and B shoots him to death. The killing of the guard was not part of the
original criminal plan. When the guard was shot, A was not there, did not
know about it, did not agree to it, and did not commit the shooting.
The legal question in this example concerns A’s criminal liability for
homicide, in addition to his obvious criminal liability for robbery. A does
not satisfy either the factual or the mental element requirements of homi-
cide, because he did not physically commit it, nor was he aware of it. The
homicide was not part of the perpetrators’ joint criminal plan. The ques-
tion may be expanded also to the inciters and accessories to robbery, if any.
In general, the question of the probable consequence liability refers to the
criminal liability of a person for unplanned offenses that were committed
by another. Before applying the probable consequence liability to human-a i
offenses, we must explore its features.106
There are two extreme opposite approaches to this general ques-
tion. The first calls for imposition of full criminal liability on all par-
ties; the second calls for broad exemption from criminal liability for any
party that does not meet the factual and mental element requirements
of the unplanned offense. The first is considered problematic for over-
criminalization, while the second is considered problematic for under-
criminalization. Consequently, moderate approaches were developed and
embraced by various legal systems.
The first extreme approach does not consider at all the factual and
mental elements of the unplanned offense. This approach originates in
Roman civil law, which has been adapted to criminal cases by several legal
systems. According to this approach, any involvement in the delinquent
event is considered to include criminal liability for any further delinquent
event derived from it (versanti in re illicita imputantur omnia quae se-
quuntur ex delicto).107 This extreme approach requires neither factual nor
mental elements for the unplanned offense from the other parties, besides
the party that actually committed the offense and accounts for both factual
and mental elements.
According to this extreme approach, the criminal liability for the un-
planned offense is an automatic derivative. The basic rationale of this ap-
proach is deterrence of potential offenders from participating in future
criminal enterprises by widening the criminal liability to include not only
the planned offenses, but the unplanned ones as well. The potential party
must realize that his personal criminal liability may not be restricted to
specific types of offenses, and that he may be criminally liable for all ex-
pected and unexpected developments that are derived directly or indi-
rectly from his conduct. Potential parties are expected to be deterred and
avoid involvement in delinquent acts.
This approach does not distinguish between various forms of involve-
ment in the delinquent event. The criminal liability for the unplanned of-
fense is imposed regardless of the role of the offender in the commission
of the planned offense as perpetrator, inciter, or accessory. The criminal
liability imposed for the unplanned offense does not depend on meeting
factual and mental element requirements by the parties. If the criminal lia-
bility for the unplanned offense is imposed on all parties of the original en-
terprise, including those who could have no control over the commission
of the unplanned offense, the deterrent value of this approach is extreme.
Prospectively, this approach educates individuals to keep away from
involvement in delinquent events, regardless of the specific role they may
potentially play in the commission of the offense. Any deviation from the
criminal plan, even if not under the direct control of the party, is basis for
criminal liability for all individuals involved, as if it had been fully per-
petrated by all parties. The effect of this extreme approach can be broad
and encompassing. Parties to another (third) offense, different from the
unplanned offense, who were not direct parties to the unplanned offense,
may be criminally liable for the unplanned offense as well, if there is the
slightest connection between the offenses. The criminal liability for the
unplanned offense is uniform for all parties and requires no factual and
mental elements. Most Western legal systems consider such a deterrent
approach too extreme, and have rejected it.108
The second extreme approach is the exact opposite of the first, and
focuses on the factual and mental elements of the unplanned offense. Ac-
cording to this approach, to impose criminal liability for the unplanned
offense, both factual and mental element requirements must be met by
each party. Only if both requirements of the unplanned offense are met by
a given party is it legitimate to impose criminal liability on it. Naturally,
because the unplanned offense was not planned, it is most unlikely that
any of the parties would be criminally liable for that offense, except for
the party that actually committed it. This extreme approach ignores the
social endangerment inherent in the criminal enterprise, which includes
not only planned offenses, but unplanned ones as well.
In light of this extreme approach, offenders have no incentive to re-
strict their involvement in the delinquent event. Prospectively, any party
that wishes to escape criminal liability for the probable consequences of
the criminal plan needs only to avoid participation in the factual aspect of
any further offense. Such offenders would tend to involve more parties in
the commission of the offense in order to increase the chance for commit-
ting further offenses, and therefore most modern legal systems prefer not
to adopt this extreme approach.
Several moderate approaches have been developed to meet the dif-
ficulties raised by these two extreme approaches. The core of the mod-
erate approaches lies in the creation of probable consequence liability,
that is, criminal liability for the unplanned offense, whose commission is
the probable consequence of the planned original offense. Probable con-
sequence refers both to mental probability from the point of view of the
party participating in the offense, and to the factual consequence resulting
from the planned offense. Thus, probable consequence liability generally
requires two principal conditions to impose criminal liability for the un-
planned offense:
1. a factual condition — the unplanned offense should be the con-

sequence of the planned offense;
2. a mental condition — the unplanned offense should be probable
(foreseeable by the relevant party) as a consequence of the com-
mission of the planned offense.
The factual condition (“consequence”) requires the incidental oc-

currence of the unplanned offense in relation to the planned one. There
should be a factual causal connection between the planned and the un-
planned offense. For example, A and B conspire to rob a bank, and ex-
ecute their plan. During the robbery, B shoots the guard to death, an act
that is incidental to the robbery and to B’s role in it. Had it not been for
the committed robbery, no homicide would have occurred. Therefore, the
homicide is the factual consequence of the robbery, and it was committed
incidentally to the robbery.
An incidental offense is one that was not part of the criminal plan, and
the parties did not conspire to commit it. If the offense is part of the crimi-
nal plan, the probable consequence liability is irrelevant, and the general
rules of complicity apply to the parties to the offense. Unplanned offenses
fall short of this requirement and create an under-criminalization problem.
The probable consequence liability is an attempt to address this difficulty
by expanding the criminal liability for unplanned offenses despite the fact
that they are unplanned.
The unplanned offense may be a different offense from the planned
one, but not necessarily. The offense may also be an additional, identical
offense. For example, A and B conspire to rob a bank by breaking into
one of its safes. A is intended to break into the safe and B to watch the
guard. They execute their plan, but in addition, B shoots and kills the
guard, and A breaks into yet another safe. The unplanned homicide is
a different offense from the planned robbery. The unplanned robbery is
identical to the planned robbery. Both unplanned offenses are inciden-
tal consequences of the planned offense, although one is different from
the planned offense and the other is identical to it. The planned offense
serves as the causal background for both unplanned offenses, as they in-
cidentally derive from it.109
The mental condition (“probable”) requires that the occurrence of the
unplanned offense be probable in the eyes of the relevant party, meaning
that it could have been foreseen and reasonably predicted. Some legal sys-
tems prefer to examine the actual and subjective foreseeability (the party
has actually and subjectively foreseen the occurrence of the unplanned
offense), whereas others prefer to evaluate the ability to foresee through
an objective standard of reasonability (the party has not actually foreseen
the occurrence of the unplanned offense, but any reasonable person in his
state could have). Actual foreseeability parallels the subjective mens rea,
whereas objective foreseeability parallels objective negligence.
For example, A and B conspire to rob a bank. A is intended to break
into the safe and B to watch the guard. They execute the plan, and B shoots
and kills the guard while A breaks into the safe. In some legal systems,
A is criminally liable for the killing only if A had actually foreseen the
homicide, and in others, if a reasonable person could have foreseen the
forthcoming homicide in these circumstances. Consequently, if the rel-
evant accomplice did not actually foresee the unplanned offense, or any
reasonable person in the same condition could not have foreseen it, he or
she is not criminally liable for the unplanned offense.
This type of approach is considered moderate because it provides an-
swers to the social endangerment problem, and at the same time it has a
positive relation with the factual and mental elements of criminal liabil-
ity. The factual and mental conditions are the starting terms and minimal
requirements for the imposition of criminal liability for the unplanned
offense.
Legal systems differ on the legal consequences of probable conse-
quence liability. The main factor in these differences is the mental condi-
tion. Some legal systems require negligence, but others require mens rea,
and the consequences may be both legally and socially different. Moder-
ate approaches that are close to the extreme approach, which holds that
all accomplices are criminally liable for the unplanned offense (versanti
in re illicita imputantur omnia quae sequuntur ex delicto), impose full
criminal liability for the unplanned offense if both factual and mental
conditions are met. According to these approaches, the party is criminally
liable for unplanned mens rea offenses even if he or she may have been
merely negligent.
More lenient moderate approaches do not impose full criminal liabil-
ity on all the parties for the unplanned offense. These approaches can
show more leniency with respect to the mental element, and match the
actual mental element of the party to the type of offense. Thus, the negli-
gent party in the unplanned offense is criminally liable for a negligence
offense, whereas the party who is aware is criminally liable for a mens rea
offense.110 For example, A, B, and C plan to commit robbery. The robbery
is executed, and C shoots and kills the guard. A foresaw this, but B did not,
although a reasonable person would have foreseen this outcome under the
circumstances.
All three are criminally liable for robbery as joint perpetrators. C is
criminally liable for murder, which is a mens rea offense. A, who acted
under mens rea, is criminally liable for manslaughter or murder, both
mens rea offenses. But B was negligent regarding the homicide, and is
therefore criminally liable for negligent homicide. Negligent offenders are
criminally liable for no more than negligence offenses, whereas other of-
fenders, who meet mens rea requirements, are criminally liable for mens
rea offenses.
American criminal law imposes full criminal liability for unplanned
offense equally on all parties of the planned offense111 as long as the un-
planned offense is the probable consequence of the planned one.112 Ap-
propriate legislation has been enacted to accept the probable consequence
liability, and has been considered constitutionally valid.113 Moreover, in
the context of homicide, American law incriminates incidental unplanned
homicide committed in the course of the commission of another planned
offense as murder, even if the mental element of the parties was not ad-
equate for murder.114
By comparison, European- Continental legal systems and English
common law impose criminal liability for the unplanned offense equally
on all parties of the planned offense.115 The English116 and European-
Continental moderate approaches are closer to the first extreme approach.117
For the applicability of the probable consequence liability to human-
i offenses, we must distinguish between two types of cases, as shown in
a
our opening examples of computer programmers at the beginning of this
section. The first type refers to cases in which the programmer designed
the ai system to commit a certain offense, but the system exceeded the pro-
grammer’s plan quantitatively (and committed more offenses of the same
type), qualitatively (committed more offenses of different types), or in both
respects. The second type refers to cases in which the programmer did not
design the ai system to commit any offense, but the system committed an
offense nevertheless.
In the first type of cases, criminal liability is divided into the liabil-
ity for the planned offense and for the unplanned one. If the program-
mer designed the system to commit a certain offense, this is perpetration-
through-another of that offense, at most. The programmer instructed the
system what it should do, therefore he used the system instrumentally for
the commission of the offense. In this case, the programmer is alone re-
sponsible for the offense, as already discussed.118 For this particular crimi-
nal liability, there is no difference between an ai system, some other com-
puterized system, a screwdriver, and any innocent human agent.
In the case of the exceeded offenses, a different approach is required.
If the ai system is strong, and is capable of computing the commission of
an additional offense, the ai system is considered criminally liable for that
offense according to the standard rules of criminal liability, as described
earlier.119 This completes the criminal liability of the ai system, and the
criminal liability of the programmer is determined according to the prob-
able consequence liability already described. Thus, if from the program-
mer’s point of view the additional offense was a probable consequence of
the planned offense, then criminal liability is imposed on the program-
mer for the unplanned offense in addition to the criminal liability for the
planned offense.
If the ai system is not considered strong, and is not capable of com-
puting the commission of the additional offense, the ai system cannot be
considered criminally liable for the additional offense. Under these condi-
tions, the ai system is considered as innocent agent. The criminal liability
for the additional offense remains the programmer’s alone, on the same
basis of probable consequence liability already described. Thus, if the ad-
ditional offense was a probable consequence of the planned offense from
the programmer’s point of view, the programmer is criminally liable for
the unplanned offense in addition to his criminal liability for the planned
offense.
In the second type of cases, the programmer has no intention to commit
any offense. From the programmer’s point of view, the occurrence of the
offense is not more than an unwilled accident. Because the programmer’s
initiative was not criminal, the probable consequence liability is inappro-
priate. The presence of a planned offense is crucial for the imposition of
probable consequence liability, as discussed, because the probable con-
sequence liability is intended to address unplanned developments of a
planned delinquent event and serve as a deterrent against participation in
delinquent activities.
When the programmer’s starting point is not delinquent, and from his
point of view the occurrence of the offense is accidental, then deterrence
is inappropriate and irrelevant. Applying such a mechanism to deal with

mistakes and accidents behind which there is no criminal intent would be
disproportional. Thus, if the ai system, which actually committed the of-
fense, is considered strong and capable of consolidating the requirements
of the offense, it may be criminally liable for that offense as a direct perpe-
trator. If not, the ai system is not criminally liable for the offense.
The programmer is at most negligent. In this type of case, the program-
mer’s criminal liability does not depend on the criminal liability of the ai
system. Regardless of whether the ai system is criminally liable, the crimi-
nal liability of the programmer for the unplanned offense is examined
separately. Because the programmer did not intend any offense to occur,
the mental element of mens rea does not apply to him, and his criminal li-
ability must be examined by the standards of negligence, so that he would
be criminally liable for negligence offenses at most. The standard of neg-
ligence is discussed in the next chapter, where we discuss the criminal
liability of ai entities for negligence offenses.
C l osin g the Opening Ex ample: Intention al Killin g R obo t

We opened this chapter with the example of a killing robot. We can now go
back and analyze the example based on the insights gained from this chap-
ter. There are two main actors whose criminal liability must be evaluated
in this homicide: the robot and its programmer. To examine the robot’s
criminal liability, we must first assess its technological capabilities. If the
ai robot has the technological capabilities of meeting the requirements of
mens rea offenses, it becomes a machina sapiens criminalis in this con-
text. If such capabilities are present, the robot’s criminal liability in this
case should be examined as if it were human.
Homicide (mens rea) offenses can be murder or manslaughter. Their
factual element requirement is identical, and it includes causing the death
of a human. Because the robot has physically caused the worker’s death,
the factual element requirement appears to be met. The mental element re-
quirement of murder requires intent (awareness and intent), whereas that
of manslaughter requires recklessness (awareness and recklessness). At
this point, careful analysis is needed. If, after analyzing the robot’s record
of computations, one of these requirements is met, it forms the criminal
liability of the robot for the given homicide offense. If not, negligence may
be examined, as discussed in the next chapter, and the robot’s criminal li-
ability is for negligent homicide. But if the robot lacks the technological
capabilities to meet the requirements of either mens rea or negligence of-
fenses, it does not form a machina sapiens criminalis in this context, and
it is not criminally liable for the homicide. The only criminal liability for
the homicide would be the programmer’s and that of other related parties
(users, the factory, and so on). Here, programmer refers not only to the
specific person or persons who actually programmed the system, but to
all related persons, including the company that manufactured the robot.
The programmer’s criminal liability is determined based on his role
in bringing about the homicide. If the programmer instrumentally used
the robot to kill the worker by designing it to operate in this way, he is
a perpetrator-through-another of the homicide. This is also the case if
the robot has no advanced ai capabilities for meeting the requirements
of the offense. In these cases, the programmer and the related parties are
alone criminally liable for the homicide, and their exact criminal liabil-
ity for the homicide, whether murder or manslaughter, is determined by
the manner in which they met the mental element requirements of these
offenses.
If the programmer designed the robot to commit an offense (any of-
fense) other than the homicide of the worker, and the ai system of the
robot continued the delinquent activity through an unplanned offense (the
homicide of the worker), then the programmer’s criminal liability for the
homicide is determined based on probable consequence liability. The ro-
bot’s criminal liability in this case is examined according to its capability
to meet the requirements of the given offense. If the programmer did not
design the robot to commit any offense, but the unplanned homicide oc-
curred, then the robot’s criminal liability is examined based on its capabil-
ity to meet the requirements for homicide, and the programmer’s criminal
liability is examined according to the standards of negligence, as we will
discuss in the next chapter.
3
ai C riminal L ia bili ty for N egligence O ffen se s
One of the common applications of ai expert systems is medical,1 generally

used for more accurate diagnosis. Patient symptoms are entered into the
system using visual scanners or other means to capture verbal data, after
which the expert system analyzes the factual data and suggests a diagno-
sis to the medical staff. Based on machine learning (inductive analysis of
cases and generalization), the expert system gains its “experience” and
learns how to distinguish among various types of symptoms, and also to
ignore unimportant symptoms. The longer the system is active, the more
accurate is its diagnosis.
This method means, among other things, that identical cases brought
to the system for diagnosis at different times might be diagnosed differ-
ently in light of the experience gained by the system between diagnoses.
The same is true of human physicians. The case of one-year-old baby was
brought to an expert ai system for diagnosis.2 The baby suffered from high
fever (104 degrees F), dehydration, and general weakness. The system di-
agnosed influenza, and recommended using analgesics and an infusion of
fluids. The medical staff acted accordingly, but after five hours the baby
died. The autopsy revealed that the baby died of a severe bacterial infec-
tion; if treated with the appropriate antibiotics, the baby could have sur-
vived. The legal question is, who is to be held criminally liable for the
baby’s death?
Negligence offenses are in common use in most countries worldwide
in a context of professional malpractice. To impose criminal liability in
these offenses, both factual and mental element requirements must be met
by the apparent offender. Consequently, the question of the criminal liabil-
ity of ai entities for negligence offenses is determined by whether they are
capable of meeting these requirements accumulatively. But even if ai enti-
ties have such capability, their liability does not function as an exemption
for humans from their own criminal liability, if humans were involved in
the commission of the offense.
84
AI C r iminal L iabili ty for Neglige nce Offe nses . 85
To explore these issues, we examine the capability of ai entities to fulfill

both the factual and mental element requirements of negligence offenses,
as well as the criminal liability of humans for ai negligence offenses.

The factual element requirement structure (actus reus) is identical for all
types of offenses: intentional, negligence, and strict liability, as noted ear-
lier.3 This structure contains one mandatory component (conduct) for all
offenses, and two optional components (circumstances and results) for
some offenses. The capability of machines to meet the factual element re-
quirement considering this structure has already been discussed,4 and the
conclusions of the discussion are same within the context of negligence
offenses.

Imposition of criminal liability for negligence offenses requires meeting
both factual and mental element requirements. The mental element re-
quirement of negligence offenses is negligence. If ai technology is capa-
ble of meeting the negligence requirement, it is possible and feasible to
impose on such technology criminal liability for negligence offenses. First,
we explore the structure of the negligence requirement, and then the ful-
fillment of the requirement by ai technology.
3.2.1. Structure of N egligence R equirement

Negligence has been recognized as a behavioral standard since ancient
times. It was already mentioned in the Laws of Eshnunna in the twenti-
eth century bc ,5 in Roman law,6 in Canonic law, and in the early English
common law.7 But in these legal systems, negligence was treated as a be-
havioral standard rather than as a type of mental element in criminal law.
The behavioral standard included dangerous behavior that fails to take
into account all relevant considerations related to the individual’s given
action. Only since the seventeenth century has negligence been consid-
ered a type of mental element in criminal law.
In 1664, the English court ruled that negligence was not adequate to
convict a defendant for manslaughter, and that at least recklessness was
required.8 This ruling created negligence as a type of mental element in
criminal law. During the nineteenth century, negligence was treated as an
exception for the general requirement of mens rea.9 Accordingly, it was
required explicitly and construed strictly. The offense needed to require

negligence explicitly, in its definition, if it were to be adequate for impos-
ing criminal liability. In the nineteenth century, negligence offenses were
still quite rare.
As transportation developed, especially with the rise of automobiles,
negligence started to be used more frequently in criminal law. Deaths
caused by traffic accidents became more common, and manslaughter was
not appropriate for these cases. A lower level of homicide was required,
and the solution was found in negligent homicide.10 But when negligence
came into common use, it caused confusion. Negligence was interpreted
as requiring unreasonable conduct, and this resulted in its being confused
with recklessness of the lower level (rashness), which required taking
unreasonable risk. That confusion produced such unnecessary terms as
“gross negligence” and “wicked negligence.”11
Many misleading rulings were issued on this basis in English law,12
until the House of Lords clarified the distinction in 2003.13 American law
developed negligence as a mental element in criminal law, parallel to the
English common law and inspired by it.14 Negligence was accepted as an
exception to mens rea during the nineteenth century, but more accurately
than in English law.15 The main distinction between recklessness and neg-
ligence involves the cognitive aspect of recklessness. Whereas reckless-
ness requires the cognitive aspect of awareness, as part of the mens rea
requirement, negligence does not.16
Both reckless and negligent offenders are required to take unreason-
able risks. However, the reckless offender is required to be aware of the
factual element components, but the negligent offender is not required
to be so aware.17 Negligence functions as an omission of awareness, and
it creates a social standard of conduct. An individual is required to take
only reasonable risks, measured objectively through the perspective of an
abstract reasonable person.18 The reasonable person is aware of his or her
factual behavior, and takes only reasonable risks.19 Naturally, it is the court
that determines how reasonable a risk is, and it does so retrospectively in
relation to a particular case.
The modern development of negligence in American law is reflected
in the American Law Institute’s Model Penal Code, inspired by modern
European-Continental understandings of negligence. Accordingly, negli-
gence is a type of mental element in criminal law, and as such it relates to
the factual element components. Negligence requires lack of awareness of
the factual element components, whereas a reasonable person could and
should have been aware of them, and requires taking unreasonable risk
regarding the results of the offense. This development has been embraced
by American courts.20
How, then, can negligence function as a mental element in criminal
law if the offender is not even required to be aware of his or her conduct?
Some scholars have asked to exclude negligence from criminal law and
leave it for tort law or other civil proceedings.21 But the justification of neg-
ligence as mental element focuses on its function as an omission of aware-
ness. Exactly the same way that act and omission are considered identical
for the imposition of criminal liability, as discussed earlier in relation to
the factual element,22 so can both awareness and omission of awareness
be considered as the basis for meeting the mental element requirement.
Following the analogy with the factual element, negligence does not
parallel inaction, but parallels omission. If a person was simply not aware
of the factual element components, and nothing more, she is not consid-
ered negligent, but innocent. Omitting to be aware means that the person
was not aware although a reasonable person could and should have been;
that is, the individual is considered to not be using her existing capabili-
ties of forming awareness. If the individual was not aware but was capable
of being aware, she also had the duty to be aware (non scire quod scire
debemus et possumus culpa est).
Negligence does not incriminate individuals who are incapable of
forming awareness, but only those who failed to use their existing capa-
bilities to do so. Negligence does not incriminate the blind person for not
seeing, but only people who have the ability to see but failed to use this
ability. Wrong decisions are a common part of daily human life, so neg-
ligence does not apply to situations of this type. Taking risks is also part
of human life, and therefore negligence is not intended to inhibit it, and
indeed society encourages taking risks in many situations. Negligence ap-
plies to taking unreasonable risks.
If people were to take no risks at all, then human development would
stop in its tracks. If we had taken no risks, we would have been staring at
the burning branch after it was struck by lightning, afraid to reach out, grab
it, and use it for our needs. Society constantly pushes people to take risks,
but reasonable ones. The question is, how can we identify the unreasonable
risks and distinguish them from the reasonable ones, which are legitimate?
For example, scientists propose to develop an advanced device that
would significantly ease our daily lives. It is comfortable, fast, elegant,
and accessible, but using it may cause the deaths of about 30,000 people
each year in the United States alone. Would developing such a device be
considered a reasonable risk?
Thirty-thousand victims a year appears to be an enormous number,

which makes using the device completely unreasonable. But using the
device in question, which we normally call a “car,”23 is not considered
unreasonable in most countries in the world today, although in the late
nineteenth century it was. The same is true for trains, airplanes, ships, and
many of our common means of transport and other everyday tools. The
degree to which a risk is reasonable is relative to its nature, and is rela-
tively determined with respect to time, place, society, culture, and other
circumstances. Different courts in different countries find different types
of people and behaviors to be reasonable within the context of negligence.
Whether a person was reasonable must be determined not only in the
abstract, but also with reference to the relevant circumstances of the spe-
cific offender. For example, it is not appropriate to compare medical deci-
sions of a physician with the behavior of an abstract reasonable person;
such decisions should be compared with those of a reasonable physician
of the same level of expertise and experience, acting in the same or similar
clinical circumstances (emergency or other), having the same resources at
his or her disposal, and so on. Such an approach would make the stan-
dard of reasonability more focused, but also subjective rather than purely
objective. A detailed discussion of this process in relation to ai systems
follows.24
Most negligence offenses are result offenses because society prefers to
use negligence to protect against factual harms resulting from unreason-
able risk taking. But negligence may be required for conduct offenses as
well. The structure of negligence is described schematically in table 3.1.
The general structure of negligence has no volitive aspect, only a
cognitive aspect. Because volition is supported by cognition, and negli-
gence does not require awareness, negligence cannot require volitive com
ponents. The cognitive aspect of negligence consists of lack of awareness
of all factual element components. The requirements of negligence in re-
lation to conduct and circumstances are identical. Both require lack of
awareness of the component (conduct/circumstances) despite the capabil-
ity to form such awareness, when a reasonable person could and should
have been aware of that component.
The reasonability in the components of negligence is examined with
reference to the capability and duty of the offender to form awareness,
although in practice, no awareness has been formed by him or her. With
respect to the results, the negligence requirement requires lack of aware-
ness of the possibility of the results occurring, despite the capability to
form such awareness, when a reasonable person could and should have
Table 3.1 . General Structure of Negligence
f ac tual elemen t componen t negligence componen t
Lack of awareness of conduct

Conduct despite the capability to form such
awareness, when a reasonable per-
son could and should have been
aware of that conduct.
Lack of awareness of circumstances
Circumstances despite the capability to form such
aware of the circumstances.
Lack of awareness of the possibil-
Results ity of the occurence of the results
despite the capability to form such
aware of that possibility as an unrea-
sonable risk.
been aware of that possibility as an unreasonable risk. For this component,

the reasonability focuses on identifying the occurrence of the result as a
possibility involving unreasonable risk. In other words, under the given
circumstances, the offender took unreasonable risks in the event at hand.
The modern structure of negligence is consistent with the concept of
criminal law as a matrix of minimal requirements, described earlier.25 It
contains both internal and external aspects. Inwardly, negligence is the
minimal mental element requirement for each of the factual element com-
ponents. Consequently, if negligence is proven in relation to circumstances
and results, but awareness is proven in relation to the conduct, this satis-
fies the requirement of negligence. This means that for each of the factual
element components, at least negligence is required, but not exclusively.
Outwardly, the mental element requirement of negligence offenses is sat-
isfied through at least negligence, but not exclusively. This means that
criminal liability for negligence offenses may be imposed by proving mens
rea as well as negligence.
Because negligence is still considered an exception to the general re-
quirement of mens rea, negligence has been required as the mental ele-
ment of relatively light offenses. In some legal systems, negligence has
even been restricted ex ante to light offenses.26 This general structure of
negligence is a template that includes terms from the mental terminology

(for example, reasonability). To explore whether ai entities are capable of
fulfilling the negligence requirement of given offenses, we must examine
these terms.
3.2.2. F ulfillment of the N egligence R equirement:

I s O bjectivity Subjective?
Sonja: What prevents you from murdering somebody?
Boris: Murder’s immoral.
Sonja: Immorality is subjective.
Boris: Yes, but subjectivity is objective.
— Diane Keaton and Woody Allen as Sonja and Boris
Grushenko in Love and Death, United Artists, 1975
The core of the negligence template in relation to the factual element com-
ponents is expressed by lack of awareness of the factual component de-
spite the capability to form such awareness, when a reasonable person
could and should have been aware of that component. Lack of awareness
is naturally the opposite of awareness, which is required by mens rea, as
discussed earlier.27 Consequently, for a person to be considered aware of
certain factual data, two accumulative conditions are required: (1) absorb-
ing the factual data through the senses; and (2) creating a relevant general
image with regard to this data in the brain. If either of these conditions is
missing, the person is not considered to be aware.
Lack of awareness may be achieved by the absence of at least one of
these two conditions. If the factual data has not been absorbed by the
senses, or if it was absorbed but no relevant general image has been cre-
ated, then the situation is considered one of lack of awareness to that fac-
tual data. Awareness is a binary situation in the context of criminal law;
therefore, no partial awareness is recognized. If portions of the awareness
process are present but the process has not been accomplished in full, the
person is considered to be unaware of the relevant factual data. This is true
for both human and ai offenders.
For lack of awareness to be considered omission of awareness, it must
occur despite the capability of the offender to form awareness, when a rea-
sonable person could and should have been aware. These are two separate
conditions: (1) possessing the cognitive capabilities needed to consolidate
awareness; and (2) a reasonable person could and should have been aware
of the factual data. The first condition has to do with the offender’s physi-
cal capabilities. If the offender lacks these capabilities, regardless of the
offense, no criminal liability for negligence may be imposed on him, just

as we cannot punish the blind for not seeing.
Therefore, negligent offenders are only those who possess the capabil-
ity of forming awareness. That is true for both human and ai offenders.
Therefore, to impose criminal liability for a negligence offense on an ai
entity, the system must possess the capability of forming awareness. An
ai system that lacks such capability cannot be held liable for an offense
of negligence, and naturally it cannot be held liable for mens rea offenses
either. These capabilities must be proven based on the general features of
the ai system, regardless of the case at hand.
If we have established that the offender was not aware of the relevant
factual data, and that he has the capability of being aware, for the offense
to become negligence we must still prove that a reasonable person could
and should have been aware of the factual data under similar circum-
stances. The “reasonable person” is a mental construct established for the
purpose of comparison. Although in some other legal spheres, the reason-
able person refers to a standard higher than that of the average person, in
criminal law it refers to the average person.28 Different societies and cul-
tures, at different times and in different places, assign different content
and qualities to the reasonable person.
The reasonable person is expected to reflect the existing situation in
the given society, and not be used by the courts to change the current situ-
ation. The standard relates to the cognitive processes that should have oc-
curred. The reasonable person is assessed based on two paths of cognitive
activity: he or she should (1) take into account all relevant considerations,
and (2) assign the proper weight to these considerations. Therefore, ig-
norance of relevant considerations is considered to be unreasonable, and
after all relevant considerations have been taken into account, they must
also be properly weighed for the action to be considered reasonable.
In general, the relevance of the considerations and their proper weight
must be determined by the court ex post. The complexity and variety
of common life situations makes it impossible to characterize a general
type of reasonable person. Such a type would be purely objective and at
the same time unrealistic and irrelevant to too many real-life situations.
Therefore, the objective standard for the reasonable person must be set
with some degree of subjectivity if it is to match any particular case. The
subjective setting is the link between the individual offender and the ob-
jective and general standard of a reasonable person.
Thus, the reasonable person must be not only assessed by the standards
of a general abstract person, but adapted to the specific circumstances of
the individual offender. We expect different people to behave differently.

Experienced attorneys act differently from inexperienced ones, even in
the same situations. The same is true for pilots, drivers, physicians, police
officers, and indeed, for all of us. Moreover, the same person acts differ-
ently under differing circumstances, weighing considerations differently.
A soldier under enemy attack, in an emergency, in a life-threatening situ-
ation, acts differently than during routine training.
Therefore, the reasonable person relevant for a given case is not a gen-
eral standard of objectivism, but it includes subjective reflections of the
individual offender.29 A surgeon with ten years of experience, acting in
an emergency situation and using the limited resources available in the
surgery, would be compared with a reasonable surgeon having the same
attributes. The attributes relevant for that comparison, as well as their con-
tent and effect, are determined by the court, which may resort to the expert
opinion of professionals for this purpose.
The reasonable person forms a sphere of reasonability that contains
all types of reasonable behaviors in the given situation. The assumption
is that there are several reasonable ways to behave in these situations,
and only deviations from the sphere of reasonability can form negligence.
When the factual data relate to the results of the conduct, the possibility
of the occurrence of the results is considered an unreasonable risk. In this
context, taking unreasonable risks is considered acting outside the sphere
of reasonability with respect to the results.30
Reasonable and unreasonable risks are assessed in the same way as
reasonable and unreasonable persons, as just described. For a risk to be
considered reasonable, the individual must take into account all relevant
considerations and assign the proper weight to them. Based on these con-
siderations, if taking the risk is one of the available options, then the risk
is considered reasonable; if not, it is considered unreasonable. To meet the
negligence requirement, an ai entity must make unreasonable decisions.
The ultimate question is whether a machine can be reasonable, or perhaps
whether a machine can be unreasonable.
From an analytical point of view, the reasonability of a machine is not
different from that of a human. Both must take into account the relevant
considerations and assign the proper weight to them. This can easily be a
matter of calculation. The relevant considerations are not more than terms
in the equation, and their proper weight is the combinations of these terms.
The equation may be constant if so programmed, but machine-learning
features can change that. Machine learning is a process of generalization
by induction from many specific cases, as we have already discussed.31
Machine learning enables the ai system to change the equation from time
to time.
Indeed, effective machine learning should cause changes in the equa-
tion almost every time a given case is analyzed. This is what happens to
our image of the world as our life experience becomes richer and broadens.
Without machine learning, the equation remains constant and the system
becomes ineffective. Expert systems that are not capable of machine learn-
ing are not different from human experts who insist on not updating their
own knowledge. Machine learning is essential for the ai system to con-
tinue developing.
When the ai system is activated for the first time, the equation and its
terms are programmed by human experts who determine what is a reason-
able course of conduct in individual cases. After analyzing a few cases,
the system begins to identify exceptions, wider and narrower definitions,
new factors and new connections between existing ones, and so on. The
ai system generalizes the knowledge absorbed from individual cases by
reformulating the relevant equations. We use the term equation to describe
the relevant algorithm, but it is not necessarily an equation in the mathe
matical sense.
Reformulating equations raises the possibility that the ai entity will
make different decisions in the future than it made in the past. This pro-
cess of induction is at the core of machine learning, and changes in the
equation produce each time a different sphere of right decisions. For ex-
ample, if on its first activation, a medical expert system is given a list of
symptoms of both common cold and influenza, it produces a diagnosis
based on those symptoms alone. But after being exposed to more cases, the
system learns to take into account other symptoms that may be essential
for distinguishing between the two conditions.
If the system is required to recommend medical treatments, these will
differ for different diagnoses. At times, the expert system will not be “sure”
because the symptoms can match more than one disease, and the system
will assess the probabilities based on measurement and analysis of the
various factors. For example, the expert system may determine that there
is a probability of 47 percent that the patient has a common cold, and of
53 percent that she has influenza. Processing these probabilities may be
the cause of negligence on the part of ai systems, and of mistaken conclu-
sions, both sure and unsure. The system may be sure in the case of a mis-
taken conclusion, and can also assess probabilities mistakenly.
Mistakes may be caused by inauspicious changes of the equation,
wrong factors being considered, ignorance of certain factors, or wrong
weights assigned to some factors. These mistakes are the by-product of

errors in the machine-learning process, or more precisely, ex post errors,
which are considered errors only after the decision has been made and
based on the consequences of that decision. Humans tend to learn empiri-
cally by trial and error. From the analytical point of view, in this context
machine and human errors are of the same type. Understanding the error,
its causes, and the ways to avoid it in the future are part of the learning
process, for both humans and machines.
The question is, what is to be considered a reasonable decision in this
context? The principal question is whether, in light of the system’s ini-
tial state with respect to the basic factual data and the experience it has
acquired through machine learning, a reasonable person could be aware
of the relevant factual data. And the derivative question is, who is this
reasonable person: a human or a machine? If we accept the concept that
objectivity in negligence offenses is subjective to some degree, as already
discussed, then the reasonable person must have attributes similar to those
of the offender. Only then can the reasonability of the offender’s decision
be assessed without causing injustice.
Therefore, if the offender is capable of machine learning, so should be
the reasonable person, according to this concept. It follows that in deter-
mining the reasonability of the ai entity’s decisions, the reasonable person
would be a reasonable ai system of the same type. It may be problematic
for the human programmers, operators, and users to escape criminal li-
ability and impose it on the mistaken machine. It would be convenient for
the medical staff to use an expert system and follow its recommendations,
then if the system is wrong, to have criminal liability imposed on it alone.
But the legal situation is not that simple.
The decisions of assigning an ai system to its position, using it, follow-
ing its recommendations, and so on are all subject to negligence offenses
as well. The ai system is capable of meeting the negligence requirements,
but this does not exempt other individuals involved in the situation from
criminal liability. The very decision to use the ai system is itself subject to
criminal liability. For example, if the decision was made with full aware-
ness of particular mistakes that the system can make, and if these mistakes
resulted in death, the human decision may lead to a charge of murder. But
if there was no such awareness, although a reasonable person in a similar
situation could and should have been aware of the risks, this may lead to
a charge of negligent homicide.
Evaluation of the reasonable machine has to do with the machine-
learning features of the system. The imposition of criminal liability in
negligence offenses is based on analysis of the machine-learning process

that led to the mistaken decision. The records maintained by the ai system
itself are the basis for the analysis. But the reasonability of the decision-
making process within the framework of machine learning may be based
on expert opinion. This is how negligence is proved in court in the case
of human offenders. It is not uncommon to prove or to refute the human
negligence using expert opinions.
For example, when the medical expert system produces probabilities
of 47 percent for common cold and 53 percent for influenza, a medical
expert may explain to the court why these probabilities are reasonable
or unreasonable for the case at hand, and a computer scientist may ex-
plain to the court the process of producing these probabilities based on the
machine’s learning process and existing database. Accordingly, the court
must ask three questions:
1. Was the a i system unaware of the factual components?

2. Does the a i system have the general capability of consolidating
awareness of the factual components?
3. Could a reasonable person have been aware of the factual com-
ponents?
If the answer is affirmative to all three questions, and this is proven

beyond any reasonable doubt, the ai system has met the requirements of
the negligence offense. ai systems that are capable of forming awareness
of mens rea offenses, as discussed earlier,32 are not limited either techno-
logically or legally in consolidating negligence for negligence offenses,
because negligence requires a lower level of mental element than mens
rea does. Negligence is therefore relevant to ai technology, and it is pos-
sible to prove it in court. But we must still answer the question of who is
criminally liable for the commission of this type of offenses.
3.3. C r iminall y L iable Entities for ai Neglige nce Offe nses

In general, imposition of criminal liability for negligence offenses requires
meeting both the factual and mental element requirements of these of-
fenses. Humans are involved in the creation, design, programming, and
operation of ai technology and entities. Therefore, when the factual and
mental elements of the offense are fulfilled by an ai entity, the question
is, who is to be criminally liable for the offenses committed? The possible
answers may be the ai entity, the related humans, or both. We will now
discuss these answers.

In negligence, as in mens rea offenses, when the offender meets both the
factual and mental element requirements of an offense, criminal liability
is imposed. As in mens rea offenses, the court is not expected to check
whether the offender was “evil,” “immoral,” or so on, as noted earlier.33
This is true for every type of offender: humans, corporations, and ai enti-
ties. Therefore, the justifications for the imposition of criminal liability
on ai entities in mens rea offenses are relevant in the case of negligence
as well. As long as these requirements are even narrowly met, criminal li-
ability can be imposed.
Negligence offenses, however, differ from mens rea offenses in their
social purpose. The pertinent question is whether their differing social
purpose is relevant not only for humans and corporations, but also for
ai entities. Initially, negligence offenses were not intended to deal with
“evil” people, but with individuals who make mistakes. Therefore, the
debate about villainy in criminal law is not relevant to negligence offenses,
as it may be for mens rea offenses. In the context of negligence offenses,
criminal law serves to help draw the boundaries of individual discretion.
Any person can make a wrong judgment, most of the time without
contradicting the norms of criminal law. For example, we may choose the
wrong spouse, car, employer, apartment, or even faith. But when mistaken
use of our discretion leads to someone’s death (negligent homicide), we do
breach a norm of criminal law.34
As long as our exercise of wrong judgment does not contradict criminal
law, society expects us to learn our lesson on our own, so that the next time
we choose a car, house, employer, and so on, we will be more careful. This
is how we gain life experience. In these cases, society takes the risk that we
may not learn our lesson, but it does not intervene through criminal law.
Yet when criminal offenses are committed, society does not take the risk
of allowing the individual to be home-schooled, because the social harm
is too serious to assume such risk.
In these cases, society intervenes by means of negligence offenses. The
purpose of these offenses is to increase the likelihood that the individual
will learn the relevant lesson. Prospectively, it is assumed that after the
lesson has been taught, the probability of committing the same offense
again is much lower. In this way, society educates physicians to be more
careful during surgery, employers to be more assiduous in protecting their
employees’ lives and welfare, construction companies to be more scrupu-
lous in observing safety regulations, factories to create less pollution, and
so on. Human and corporate offenders are expected to learn their lessons
through the intermediary of criminal law. Should this type of education

be relevant to ai entities as well?
The answer is affirmative. For the educational purpose of the criminal
negligence offense, it is true that there is not much use in the imposition
of criminal liability unless the offender has the ability to learn. If we want
the offender to learn a lesson after making a mistake, we must assume that
the offender is capable of learning. If such capability is present and exer-
cised, then criminal liability for negligence offenses is necessary. But if no
such capability is exercised, then there is no prospective value in impos-
ing criminal liability for negligence because the results would be the same,
regardless of whether criminal liability is imposed.
For ai systems, which are equipped with the capabilities necessary for
machine learning, criminal liability for negligence offenses is necessary.
As for humans, negligence offenses can draw the boundaries of discretion
of ai systems. Humans, corporations, and ai systems alike are expected to
learn from their mistakes and improve their decisions prospectively. When
the mistakes fall under its purview, criminal law intervenes in shaping
the decision maker’s discretion. For ai systems, criminal liability for neg-
ligence offenses is a chance to reconsider the decision-making process in
light of external limitations dictated by criminal law.
We have learned over the years that improving the human process of
decision making requires criminal liability for negligence, and the same
logic applies to ai systems equipped with machine-learning functions.
Although ai systems can be reprogrammed, their precious experience,
gained through machine learning, would be lost. ai systems can learn the
limits of their discretion on their own, but this is true for humans as well
(and we must still impose criminal liability on humans for negligence
offenses).
In sum, if ai entities are capable of fulfilling both the factual and mental
element requirements of criminal liability for negligence offenses, and if
the rationale for the imposition of criminal liability for these offenses is
relevant to both humans and ai systems, there is no reason to avoid im-
posing criminal liability in these cases. But this is not the only type of
ai involvement in criminal liability for negligence offenses.

and Semi-Innocent A gents
As described earlier, the most common way to deal with the instrumen-
tal use of individuals in the commission of offenses is the general form
of perpetration-through-another.35 To impose criminal liability for perpe-
tration of an offense through another, it is necessary to prove awareness

of such instrumental use. Therefore, perpetration-through-another is ap-
plicable only in the case of mens rea offenses. In most cases, the person
being instrumentally used by the perpetrator is considered an innocent
agent, and no criminal liability is imposed on him or her. The analysis
of perpetration-through-another in the context of mens rea offenses has
already been discussed.
Nevertheless, a person instrumentally used can also be considered a
semi-innocent agent who is criminally liable for negligence, although at
the same time, the perpetrator is criminally liable for a mens rea offense.
This is the case when negligence may be relevant for the perpetration-
through-another, and that completes the discussion toward it. For exam-
ple, a nurse in an operating room realizes that a person who had attacked
her in the past is about to be operated on, and she decides that he deserves
to die. She infects the surgical instruments with lethal bacteria, telling the
surgeon that the instruments have been sterilized.
A few hours after the surgery, the patient dies as a result of an infec-
tion. Legal analysis of the case suggests that the nurse is the perpetrator-
through-another of murder, having instrumentally used the surgeon to
commit the murder. The surgeon’s criminal liability in this case depends
on his mental state. If he is an innocent agent, he is exempt from criminal
liability. But if he has the legal duty to make sure that the instruments have
been sterilized, he is not an entirely innocent agent because he failed to
fulfill his legal duties.
At the same time, because he was not aware of the infected instru-
ments, this is a case of negligence. When the agent is not aware of crucial
elements of the offense, but a reasonable person in the same situation
could and should have been aware, the agent is negligent and is called a
semi-innocent agent.36 Thus, when a person instrumentally uses another
person who is negligent regarding the commission of the offense, this is
perpetration-through-another, but both persons are criminally liable: the
perpetrator for a mens rea offense (for example, murder) and the other
person for a negligence offense (for example, negligent homicide). Given
that ai systems are capable of forming negligence as a mental element, the
question is whether they can function as semi-innocent agents.
A case in which an ai system can be a semi-innocent agent is one in
which the perpetrator (human, corporation, or ai entity) instrumentally
uses an ai system for the commission of an offense, and although it was
used instrumentally, the ai system was negligent with regard to commit-
ting the offense. Only ai systems that are capable of fulfilling the mental
element requirement of negligence offenses can be considered as semi-

innocent agents and can function as such. But not in every case in which
the ai system has the capability for negligence would it automatically func-
tion as a semi-innocent agent. The capability for negligence is necessary
for functioning as a semi-innocent agent, but it is not sufficient.
The semi-innocent agent, whether human, corporation, or machine,
must be examined ad hoc in any given case. Only if the agent was neg-
ligent regarding the commission of the offense can it be considered a
semi-innocent agent. Thus, if the instrumentally used ai system did not
consolidate awareness of the relevant factual data but had the capability
to do so, and a reasonable person would have consolidated such aware-
ness, the ai system is considered a semi-innocent agent in the context of a
perpetration-through-another (itself).
The perpetrator’s criminal liability is not affected by the agent’s crim-
inal liability, if any. The criminal liability of the perpetrator-through-
another is for the mens rea offense, whether the instrumentally used
ai system has no criminal liability (that is, is an innocent agent or lacks
the relevant capabilities) or is criminally liable for negligence (that is, is a
semi-innocent agent). Thus, using the legal construction of perpetration-
through-another based on the instrumental use of an ai entity does not
affect the status of the perpetrator as a mens rea offender, regardless of the
ai system’s criminal liability.
By the same token, the agent’s criminal liability in these cases is not
directly affected by the perpetrator’s criminal liability. If the ai system
was negligent (that is, it fulfilled both factual and mental element require-
ments of the negligence offense), then criminal liability for negligence is
imposed on it, and the system is classified as a semi-innocent agent in a
perpetration-through-another. If the ai system was not negligent because
of its lack of capability or for any other reason, then no criminal liability
is imposed on it, and the system is classified as an innocent agent in a
perpetration-through-another.
In other words, if the ai system is neither an innocent agent nor a
semi-innocent agent, this means that it has fully met the requirements
of the mens rea offenses. This is not the case for perpetration-through-
another, but for principal perpetration on the part of the ai system. If the
ai system is capable of being criminally liable for mens rea offenses as a
sole offender, then there is nothing to prevent it from committing the of-
fense jointly, with other entities, humans, corporations, or other ai entities.
Complicity by the ai systems requires at least mens rea, not negligence,
because it requires awareness of the complicity and of the delinquent as-
sociation. This situation is not substantively different from meeting the

requirements of any other mens rea offense, as discussed earlier.37

Probable consequence liability deals with the commission of an unplanned
offense, different from or additional to the main, planned offense by more
than one perpetrator. The question in these cases involves the criminal
liability of the other parties to the unplanned offense committed by one
of the parties. We have already discussed probable consequence liability
for an unplanned mens rea offense.38 The relevant question here concerns
probable consequence liability to an unplanned negligence offense com-
mitted by an ai system. Are the programmers, users, and other related
persons criminally liable for unplanned negligence offense committed by
an ai system?
This question winds up our two previous discussions regarding
human criminal liability for negligence by ai entities in addition to or
instead of the criminal liability of the ai system for that offense.39 For ex-
ample, a medical expert system is used to diagnose certain types of dis-
eases by analyzing patient symptoms. The analysis is based on machine
learning and inductive generalization of various cases. When the system
fails to diagnose a case correctly, this results in the wrong treatment, lead-
ing eventually to the patient’s death. Analysis of the system’s activity re-
veals negligence by it, which meets both the factual and mental element
requirements of negligent homicide. At this point a question arises about
the programmer’s criminal liability for the offense.
The programmer’s criminal liability has nothing to do with the deci-
sion to use the ai system, to follow its diagnosis, and so on, but involves
the initial programming of the system. If the programmer had programmed
the system to kill patients and had instrumentally used it for this purpose,
this would have been perpetration-through-another of murder, but this is
not the case here. In this case, the probable consequence liability may be
relevant for the programmer’s criminal liability.
The mental condition for the probable consequence liability requires
the unplanned offense to be “probable” from the point of view of the party
that did not actually commit it. It is necessary for that party to have been
able to foresee and reasonably predict the commission of the offense.
Some legal systems prefer to examine actual and subjective foreseeabil-
ity (the party has actually and subjectively foreseen the occurrence of the
unplanned offense), whereas others prefer to evaluate the ability to fore-
see through an objective standard of reasonability (the party has not actu-
ally foreseen the occurrence of the unplanned offense, but any reasonable
person in this situation could have).
Actual foreseeability parallels subjective mens rea, whereas objec-
tive foreseeability parallels objective negligence. Consequently, in legal
systems that require objective foreseeability, the programmer should be
at least negligent for the commission of the negligence offense by the
ai system for the imposition of criminal liability for that offense. But in
legal systems that require subjective foreseeability, the programmer must
be at least aware of the possibility of the commission of the negligence of-
fense by the ai system for imposition of criminal liability for that offense.
However, if neither subjective nor objective foreseeability of the com-
mission of the offense apply to the programmer, the probable consequence
liability is irrelevant. In this case, no criminal liability can be imposed on
the programmer, and the criminal liability of the ai system for the negli-
gence offense does not affect the programmer’s liability.
C l osin g the Opening Ex ample: Neglige nt Killin g R obo t

We opened this chapter with an example describing an expert system that
killed through negligence. We can now return to this case and analyze it
based on insights achieved in this chapter. There are four main charac-
ters whose criminal liability can be considered for this homicide: the ai
system, the programmer, the hospital, and the user of the ai system. To
examine the criminal liability of the ai system, we must examine first its
technological capabilities. If the ai system has the technological capabili-
ties to meet the requirements of negligence offenses, it forms a machina
sapiens criminalis in this context.
If these capabilities are present, the ai system’s criminal liability
should be examined as if it were human. But if the ai system does not
have the technological capabilities to meet the requirements of negligence
offenses, it cannot be considered a machina sapiens criminalis and is not
criminally liable for the negligent homicide. In the opening example, the
medical expert system met the factual element requirement for negligent
homicide because its diagnosis and recommendations set off a chain of
events that ended with the baby’s death.40
To meet the mental element requirement, it is necessary to prove neg-
ligence on the part of the system. To do so, it is required to show that
the system has the capability to consolidate awareness, that it did not in
practice form awareness of the relevant factors involved in the mistaken
diagnosis, and that a reasonable person could have formed such aware-
ness. The first condition can be proven based on the characteristics of the
system and according to the definition of awareness in criminal law (per-
ception by the senses of factual data and its understanding), as discussed
earlier.41 If the system has such capability, and if that capability has been
activated, then the system formed awareness.
If the system had been aware of the bacterial infection and nevertheless
produced an inappropriate diagnosis and recommendations, this would
not have been negligence but mens rea, and the homicide offense would
have been manslaughter or murder. Otherwise, the system meets the
second condition for negligence. For the ai system to be criminally liable
for negligent homicide, its mental state must be compared with that of a
reasonable person. The subjective incarnation of the reasonable person in
this case is a reasonable medical expert system. The question is whether
the system took into consideration all the relevant factors required for
proper diagnosis, and whether they were appropriately weighed, given the
relevant circumstances of the case.
The diagnosis produced by the system did not necessarily have to be
correct for the system to be exempt from liability for negligence. Not every
incorrect diagnosis is considered negligent. At issue here is the system’s
judgment. If it is proven that the system’s judgment was correct (the cal-
culation took all factors into consideration and weighed them properly),
then the system acted reasonably. Although the diagnosis did not cor-
rectly identify the cause of the symptoms, this is still not in the realm of
negligence if the system’s judgment was reasonable. Not every mistake is
considered negligent. In this case, the court would have to examine cir-
cumstances carefully, with or without the help of medical and software
experts.
If the court reaches the conclusion that a reasonable person would
have been aware of the possibility that the symptoms were caused by a
bacterial infection, or would have considered the probability to be low that
influenza had caused the symptoms, then the system is considered negli-
gent. Otherwise, the system is exempt from criminal liability for the death
of the baby. Three other actors may be criminally liable in this case, in
addition to (but not instead of) the ai system. The criminal liability of the
ai system may affect that the liability of these actors, but not necessarily.
The programmer’s criminal liability is determined by his role in the oc-
currence of the death. If the programmer instrumentally used the ai system
to kill the baby by designing it in such a way as to produce erroneous
medical diagnoses, the programmer is a perpetrator-through-another of
the homicide. This would also be the case when the ai system had no ad-
vanced ai capabilities to meet the requirements of the offense, so that the
programmer and the other related parties, but not the ai system, are the
only ones criminally liable for the homicide. Their exact criminal liabil-
ity for the homicide, whether murder or manslaughter, is determined by
whether they meet the mental element of these offenses.
If the programmer has foreseen the consequence (the patient’s death),
or if any reasonable programmer could have foreseen it, and if the ai system
is criminally liable for the homicide, the programmer may be criminally
liable for negligent homicide through probable consequence liability, if the
offense was unplanned. Generally, if the programmer did not intention-
ally design the system to commit any offense, and an unplanned homi-
cide (murder, manslaughter, or negligent homicide) occurred, the system’s
criminal liability is examined based on its capabilities and on meeting the
requirements of the relevant homicide, and the programmer’s criminal li-
ability is examined based on the standards of negligence. Here, program-
mer refers not only to the person or persons who actually programmed the
system, but to all related persons, including the company that produced
the system.
Two other actors who may be criminally liable are the users of the
system (the medical staff) and the hospital. The hospital may be criminally
liable for making the decision to use an ai system in real-life medical cases.
To determine the liability of the hospital, it is necessary to examine the
factual data about the system that was available and known to the hospital,
including the probability that the system would make mistakes. If the deci-
sion was reasonable, the hospital is not negligent. Similarly, the medical
staff must be examined to determine its criminal liability for homicide.
The factual data about the system that was available and known to the staff
is extremely important in evaluating the decision of the medical staff to
follow the system’s diagnosis and recommendations.
If the ai system was correct in the previous 1,000 cases, then the medi-
cal staff’s decision to follow its recommendations may be reasonable. But
it is up to the court to determine the sphere of reasonability in each case.
Perhaps in emergency rooms a 10 percent rate of mistakes is reasonable,
whereas in cosmetic surgery such a rate is far beyond reasonability. Thus,
criminal liability of ai systems for negligence offenses is possible, and it
does not necessarily reduce the criminal liability of other relevant entities.
4
ai C riminal L ia bili ty for
Stric t L ia bili ty O ffen se s
The operations carried out by drones are based on artificial intelligence

technology. ai drones operate not only on the ground (unmanned vehi-
cles), but also in the air and underwater.1 For example, in 2012 the US
Navy examined an operative drone that was slated to land on an aircraft
carrier, relying on pinpoint cpr coordinates and advanced avionics. Com-
puters on the carrier digitally transmit the relevant data (speed, cross-
winds, and so on) to the drone, as it approaches for the landing. The drone
not only knows how to land by itself, but it also knows the weapons it is
carrying; if, when, and where it needs to refuel from an aerial tanker; and
whether there is a nearby threat. It does its own calculations and decides
what to do next.2
The same technology is implemented in other devices as well, with
the possible result that these devices can commit traffic offenses, whether
in the air or on the ground. For example, an unmanned vehicle is used for
ground transportation on roads travelled by other, human-driven vehicles.
When the unmanned vehicle arrives at an intersection, it analyzes the
traffic situation and decides to proceed. The vehicle has a green light, and
the lights in all other directions are red, but an ambulance arrives at high
speed and crosses the intersection.
The vehicle hits the ambulance, causing severe injury to the passen-
gers, and the patient in the ambulance dies as a result. Analysis of the pa-
tient’s medical situation shows that the patient would have survived had
the ambulance not been hit by the unmanned vehicle, so the collision is
the direct factual cause of the patient’s death. The legal question is, who is
to be held criminally liable for the patient’s death?
Strict liability offenses are in common use by most countries in a con-
text of public order and welfare (for example, traffic offenses, labor safety
offenses, or environmental offenses). For imposition of criminal liability
in these offenses, both the factual and mental element requirements must
104
AI L iabili ty for Stric t L iabili ty Offe nses . 105
be met by the apparent offender. Consequently, the question regarding

the criminal liability of ai entities for strict liability offenses is whether
they are capable of meeting these requirements accumulatively. Moreover,
even if these entities have such capability, this does not serve to exempt
humans from criminal liability if they were involved in the commission
of the offense.
To explore these issues, we examine the capability of ai entities to
meet both the factual and mental element requirements of strict liability
offenses, and the criminal liability of human offenders for strict liability
offenses committed by ai entities.

The factual element requirement structure (actus reus) is identical for
all types of offenses: intentional, negligence, and strict liability, as noted
earlier.3 This structure contains one mandatory component (conduct) for
all offenses, and two optional components (circumstances and results)
for some offenses. The capability of machines to meet the factual element
requirement, considering this structure, has already been discussed,4
and the conclusions of that discussion hold true for negligence offenses
as well.

Imposition of criminal liability for strict liability offenses requires meet-
ing both factual and mental elements requirements. The mental element
requirement of strict liability offenses is strict liability or presumed neg-
ligence. If ai technology is capable of meeting the strict liability require-
ment, it is possible and feasible to impose criminal liability on it for strict
liability offenses. We begin by exploring the structure of strict liability
requirements, and then proceed to examine the way in which ai technol-
ogy meets the requirement.
4.2.1. Structure of Strict L iability R equirement

Strict liability evolved from absolute liability and was accepted in criminal
law as a form of mental element requirement. Beginning in the eighteenth
century, English common law determined that some offenses require nei-
ther mens rea nor negligence. These offenses, called public welfare of-
fenses, were inspired by tort law, which accepted absolute liability as legit-
imate.5 Consequently, these became criminal offenses of absolute liability,
and imposition of criminal liability for them required proof of the factual
element alone.6
Absolute liability offenses were considered exceptional because they
require no mental element. In some cases Parliament intervened and re-
quired a mental element,7 and in other cases mental element requirements
were added by court rulings.8 By the mid-nineteenth century, English
courts began to consider efficiency criteria as part of criminal law in vari-
ous contexts, which gave rise to convictions on the basis of public incon-
venience. Offenders were indicted and convicted despite the fact that no
mental element was proven because of the public inconvenience caused
by the commission of the offense.9
These convictions created, in practice, an upper threshold for neg-
ligence, a type of augmented negligence, which required individuals to
follow strict guidelines and make sure they committed no offenses. This
standard of behavior is higher than the one imposed by negligence, which
requires only reasonable conduct. Strict liability offenses require more
than reasonability, making sure that no offense is committed at all. These
offenses show a clear preference of the public welfare over the strict jus-
tice for the potential offender. Because these offenses were not considered
severe, they were expanded “for the good of all.”10
This development was considered necessary because of the legal and
social changes at the time of the first industrial revolution. For example,
the increasing number of workers in the cities caused employers to de-
grade the workers’ social conditions. Parliament intervened through social
welfare legislation, and the efficient enforcement of this legislation was
through absolute liability offenses.11 It was immaterial whether the em-
ployer knew what the proper social conditions for workers were; employ-
ers had to make sure that no violation of these conditions occurred.12 In the
twentieth century, this type of criminal liability spread to other spheres of
law as well, including traffic law.13
American criminal law accepted absolute liability as a basis for crimi-
nal liability in the mid-nineteenth century,14 ignoring previous rulings that
rejected it.15 Acceptance was restricted to petty offenses, and violations
were punished by fines, and not very severe ones at that. Similar accep-
tance occurred at about the same time in the European-Continental legal
systems, so that absolute liability in criminal law became a global phe-
nomenon.16 In the meantime, the fault element in criminal law became
much more important because of internal developments in criminal law,
and mens rea became the major and dominant requirement for mental ele-
ment in criminal law.
Therefore, criminal law was required to make changes in absolute li-

ability in order to meet the modern understanding of fault, which triggered
the move from absolute to strict liability. At the heart of the change was
the move from absolute legal presumption (praesumptio juris et de jure) to
relative legal presumption (praesumptio juris tantum), so that the offender
had the opportunity to refute the criminal liability. The presumption was
presumption of negligence, either refutable or not.17 The move from abso-
lute liability to strict liability eased the acceptance of the presumed negli-
gence as yet another, third form of mental element in criminal law.
Since the broad acceptance of strict liability worldwide, legal systems
have justified it both from the perspective of fault in criminal law18 and
constitutionally. The European Court of Human Rights justified using strict
liability in criminal law in 1998,19 and thus strict liability was considered
not to be contradicting the presumption of innocence, protected by the
1950 European Human Rights Covenant.20 This ruling has been embraced
in Europe and in Britain.21 The Supreme Court of the United States ruled
consistently that strict liability does not contradict the US Constitution.22
So did the supreme courts in various states.23 But in the United States, the
courts recommended to restrict the use of these offenses to the minimum
necessary, and to prefer using mens rea or negligence offenses.
The strict liability construct in criminal law focuses on the relative
negligence presumption and the ways to refute it. According to this pre-
sumption, if all components of the offense’s factual element requirement
are proven, it is presumed that the offender was at least negligent. Conse-
quently, for the imposition of criminal liability in strict liability offenses,
the prosecution does not have to prove the mental state of the defendant,
only the fulfillment of the factual element requirements. The mental state
of the offender is derived from the conduct. Up to this point, strict liability
is similar to absolute liability.
But contrary to absolute liability, strict liability can be refuted by the
defendant because it is based on a relative legal presumption. To refute
strict liability, the offender must accumulatively meet two conditions:
(1) no mens rea or negligence existed on the part of the offender, and
(2) all reasonable measures to prevent the offense were taken.
The first condition refers to the mental state of the offender. According
to the presumption, the commission of the factual element presumes that
the offender was at least negligent. This means that the offender’s mental
state was one of negligence or of mens rea. The offender must therefore
refute the conclusion of the presumption, and prove it incorrect in his
case. To do so, the offender must prove that he was not aware to the rele
vant facts, and that no other reasonable person could have been aware of
them under the same circumstances. This proof resembles refuting mens
rea in mens rea offenses and negligence in negligence offenses.
However, strict liability offenses are not mens rea or negligence of-
fenses, because refuting mens rea and negligence is not sufficient to pre-
vent imposition of criminal liability. The social and behavioral purpose
of these offenses is to ensure that individuals conduct themselves strictly
according to rules, and commit no offenses. That should be proven as well.
The offender must therefore prove that he or she took all reasonable mea-
sures to prevent the offense.24 Note the difference between strict liability
and negligence: to refute negligence, it is sufficient to prove that the of-
fender acted reasonably, but to refute strict liability, it is necessary to prove
that all reasonable measure were taken.25
To refute the negligence presumption of strict liability, the defendant
must positively meet both conditions by a preponderance of the evidence,
as in civil law cases. The defendant is not required to prove that these
conditions have been met beyond any reasonable doubt, but in general, it
is not sufficient to merely raise a reasonable doubt. The burden of proof
is higher than the general burden of the defendant. The possibility for the
offender to refute the presumption becomes part of the strict liability re-
quirement because it relates to the offender’s mental state. The structure
of strict liability is described schematically in table 4.1.
The modern structure of strict liability is consistent with the concept
of criminal law as a matrix of minimal requirements, described earlier,26
and it has both internal and external aspects. Internally, strict liability is
the minimal mental element requirement for each of the factual element
components. Consequently, if strict liability is proven in relation to cir-
cumstances and results but negligence is proven in relation to conduct,
that satisfies the requirement of strict liability. This means that for each
of the factual element components, at least strict liability is required, but
not exclusively strict liability. Externally, the mental element require-
ment of strict liability offenses is satisfied by at least strict liability, but
not exclusively. This means that criminal liability for strict liability of-
fenses may be imposed by proving mens rea and negligence as well as
strict liability.
Because strict liability is still considered an exception to the general
requirement of mens rea, it is an adequate mental element for relatively
light offenses. In some legal systems, strict liability has been restricted ex
ante or ex post to light offenses.27 This general structure of strict liability
is a template that contains terms derived from the mental terminology. To
Table 4.1 . General Structure of Strict Liability
f ac tual elemen t componen t stric t lia bili ty componen t
Conduct The conduct has been committed,

and the offender fails to prove that
neither mens rea nor negligence
requirements for the conduct have
been met, and that all reasonable
measures to prevent the offense were
taken.
Circumstances The circumstances are present,
and the offender fails to prove that
neither mens rea nor negligence
requirements for the conduct have
been met, and that all reasonable
measures to prevent the offense were
taken.
Results The results have occurred, and the
offender fails to prove that neither
mens rea nor negligence require-
ments for the conduct have been
met, and that all reasonable mea-
sures to prevent the offense were
taken.
explore whether ai entities are capable of meeting the strict liability re-
quirement of given offenses, we must examine these terms.
4.2.2. F ulfillment of Strict L iability:

M aking the F actual Become M ental
To prove that the defendant met the requirement for strict liability, the
prosecution may choose to prove only the factual element of an offense.
According to the negligence presumption, the presence of the factual ele-
ment indicates that the offender has been at least negligent. But as already
noted,28 the possibility of refuting that presumption by the defendant is
an integral part of the substance and structure of strict liability. There-
fore, strict liability may apply only to offenders who possess the mental
capability of refuting the presumption, which has to do with the mental
element.
The mental capability of refuting the presumption does not necessarily
mean that the offender has proof of his or her innocence, or has convinc-
ing arguments to that effect, but only that the inner capabilities required
to refute the negligence presumption are present. Refuting the presump-

tion requires proof that the offender (1) was neither aware nor negligent
regarding the factual element components, and (2) has taken all reasonable
measures to prevent the commission of the offense. The required capabili-
ties, therefore, are those needed to consolidate awareness and negligence,
and those needed to act reasonably.
The question is whether ai systems possess such capabilities, and
therefore whether criminal liability for strict liability offenses may be
imposed on them. We examine these capabilities one by one. First is the
capability to consolidate awareness and negligence. We have examined
this capability of ai systems already in relation to mens rea and negli-
gence.29 ai systems that possess the relevant features discussed earlier do
indeed have this capability. Consequently, all ai systems that can be in-
dicted for mens rea and negligence offenses can also be indicted for strict
liability offenses.
This makes sense. Analogously, if mens rea and negligence require
higher mental capabilities than does strict liability, then the lower capa-
bility required for strict liability would be much easier to achieve than the
higher capabilities needed for mens rea and negligence. Therefore, the
ai offender is required to possess the same features, regardless of the
type of offense. Because negligence requires the capability to consolidate
awareness,30 and because strict liability requires the capabilities to consol-
idate awareness and negligence, it is clear that all ai offenders are required
to possess the capability of consolidating awareness.
So in this context, an indictable offender is one who has the capability
of consolidating awareness, regardless of whether this capability has been
realized and utilized, and regardless of the type of offense. An ai system
that is indictable for strict liability offenses must have the mental features
and capabilities needed to be indicted for mens rea or negligence offenses.
This fact reveals the character of the ai offender, with respect to inner ca-
pabilities, as more or less uniform. Thus, the minimal inner features and
capabilities of machina sapiens criminalis are uniform and can be defined
accordingly.
The second required capability is reasonability. Refuting the negli-
gence presumption requires proof of having taken all reasonable measures,
as already noted, and the capability of acting reasonably. Without such ca-
pability, no ai system can take all reasonable measures. The capability for
reasonability is required in negligence as well. This is the same capability
as the one required for strict liability, and we have discussed it already in
the context of negligence.31 The fact that the same capability is required
strengthens the argument just made that all inner capabilities of machina
sapiens criminalis are uniform, regardless of the type of offense.
Although the capability is the same in both negligence and strict li-
ability offenses, it operates differently in the cases of negligence and strict
liability. For negligence offenses, it is required that the ai system identify
the reasonable options and make a choice between them. The practical re-
quirement is that its final choice be a reasonable option. For strict liability
offenses, it is also required that the ai system identify the reasonable op-
tions, but in addition, it must choose to exercise all of them. The choice
here is between reasonable and unreasonable, whereas in the case of neg-
ligence, the choice is among the reasonable options.
It is much easier for an ai system, under these conditions, to act reason-
ably in a strict liability context than in the context of negligence because
in strict liability situations, fewer choices are required. Accordingly, in
relation to the criminal liability of an ai system in strict liability offenses,
the court must answer the following three questions:
1. Was the factual element of the offense fulfilled by the a i system?

2. Does the a i system have the general capability of consolidating
awareness or negligence?
3. Does the a i system have the general capability of reasonability?
If the answer to all three questions is affirmative, and this is proven beyond
any reasonable doubt, then the ai system has fulfilled the requirements of
the strict liability offense, and the ai system is presumed to be at least
negligent.
At this point, the defense has the opportunity to refute the negligence
presumption through positive evidence. After the evidence is presented,
the court must decide in two questions:
1. Has the a i system indeed formed mens rea or negligence re-

garding the factual element components of the strict liability
offense?
2. Has the a i system not taken all reasonable measures to prevent
the actual commission of the offense?
If the answer to even one of these two questions is affirmative, the neg-
ligence presumption has not been refuted, and criminal liability for the
strict liability offense is imposed. Only if the answers to both questions are
negative is the negligence presumption refuted, and no criminal liability
for the strict liability offense can be imposed on the ai system. In general,
ai systems that are capable of forming awareness and negligence, as dis-
cussed earlier,32 have neither technological nor legal limitations in forming

the inner requirements for strict liability offenses because strict liability
requires a lower level of mental element than do mens rea or negligence.
Given that strict liability is relevant to ai technology and that it is possible
to prove it in court, then who is to be criminally liable for the commission
of this type of offenses?
4.3. C ri minall y L iable Entities for ai Str ic t L iabili ty Offe nses

In general, imposition of criminal liability for strict liability offenses re-
quires meeting both factual and mental elements of these offenses. Humans
are involved in the creation, design, programming, and operation of
ai technology and entities. Consequently, when the factual and mental ele-
ments of the offense are met by the ai entity, the question is, who is to be
criminally liable for the offenses committed? The possible answers are the
ai entity, the humans, or both. We will now discuss these answers.

In strict liability offenses, such as mens rea and negligence, when the of-
fender fulfills both the factual and mental element requirements of the
offense, criminal liability is imposed. As in the case of mens rea and neg-
ligence offenses, the court is not supposed to check whether the offender
had been “evil” or “immoral,” as already discussed.33 This is true for all
types of offenders, including humans, corporations, and ai entities. There-
fore, the justifications used to impose criminal liability on ai entities in
mens rea and negligence offenses are relevant for strict liability as well.
As long as these requirements are narrowly met, criminal liability can
be imposed.
But strict liability and negligence offenses are different from mens rea
offenses in their social purpose, as we have discussed.34 The relevant ques-
tion is whether their different social purpose applies not only to humans
and corporations, but to ai entities as well. From their inception, strict
liability offenses were not designed to deal with evil people, but with
individuals who did not make all possible efforts to prevent the commis-
sion of an offense. Therefore, the debate about evil in criminal law is not
relevant to strict liability offenses in the same way that it may be relevant
for mens rea offenses.
In this context, criminal law function is to educate and ensure that no
offense is committed, and strict liability offenses serve to draw the bound-
aries of individual discretion. Any person can use wrong judgment and
make some, but not all possible efforts to prevent the commission of an
offense. Most of the time, this does not contradict the norms of criminal
law. For example, we may take the risk of investing our money in doubtful
stocks, and we may not take all possible precautions to prevent damage to
our investments, but that does not breach any criminal law.
In some cases, however, wrong judgment contradicts a norm of crimi-
nal law that was designed to educate individuals to make sure that no
offense is committed. As long as our wrong judgment and lack of effort
to prevent offenses do not contradict the criminal law, society expects us
to learn our lesson on our own. Next time we invest our money, we will
examine the details of our investment more carefully. This is how we gain
life experience. In this case, society takes a risk that we will not learn our
lesson, but still does not intervene through criminal law.
But when criminal offenses are committed, society does not assume
the risk of allowing individuals to learn their lessons on their own. In
these cases the social harm is too grave to allow such risk, and society
intervenes using both strict liability and negligence offenses. In the case
of strict liability offenses, the purpose is to educate individuals to make
every possible effort to prevent the occurrence of the offense. For example,
drivers are expected (and educated) to drive carefully to prevent any pos-
sible moving violation.
The objective of the offense is to make sure that individuals learn how
to behave extra carefully. Prospectively, it is assumed that after some-
one has been convicted of a strict liability offense, the probability for re-
commission of the offense is much lower. For example, society educates
its drivers to drive extra carefully, and its employers to adhere to state reg-
ulations regarding the payment of wages, and so on. Human and corporate
offenders are expected to learn how to behave carefully by means of the
criminal law. Is this method of education relevant for ai entities as well?
The answer is affirmative. For the educational purpose of strict liabil-
ity, there is not much use in imposing criminal liability unless the offender
has the ability to learn and change his or her behavior accordingly. If we
want to make offenders learn to behave very carefully, we must assume
that they are capable of learning and implementing that knowledge. If such
capabilities are present and exercised, it is necessary to impose criminal
liability for strict liability offenses. But if no such capabilities are exer-
cised, then imposing criminal liability is entirely unnecessary because no
prospective value is expected, and the result of using or not using criminal
liability for strict liability offenses would be the same.
For ai systems that are equipped with the relevant capabilities for ma-
chine learning, criminal liability for strict liability offenses is necessary

if these capabilities are to be applied in the relevant situations in which
there is an obligation to behave extra carefully. Exactly as for humans,
strict liability offenses may trace the boundaries of discretion for ai sys-
tems. Humans, corporations, and ai systems alike are supposed to learn
from their experience and improve their decisions prospectively, includ-
ing the standards of carefulness.
When the absence of carefulness is part of criminal law, the law inter-
venes to shape the judgment exercised with respect to careful behavior.
For the ai system, the criminal liability for strict liability offenses is an
opportunity to reconsider the decision-making process in light of external
limitations dictated by criminal law, which require extra-careful conduct
and decision making. We have learned over many generations that the
human decision-making process requires criminal liability for strict liabil-
ity in order to be improved, and the same logic applies to ai systems using
machine learning methods.
Naturally, ai systems can simply be reprogrammed each time, but this
is different from the precious experience gained through machine learn-
ing. ai systems may be capable of learning their boundaries and correcting
their judgment on their own, but the same is true for humans, and we still
impose criminal liability on humans for strict liability offenses. Conse-
quently, if ai entities have the required capabilities for meeting both fac-
tual and mental elements of criminal liability for strict liability offenses,
and if the rationale for the imposition of criminal liability for these of-
fenses is relevant to both humans and ai systems, then there is no reason
to avoid imposing criminal liability in all these cases. But this is not the
only type of ai involvement in criminal liability for strict liability offenses.

As described earlier,35 the most common way of handling the instrumen-
tal use of individuals for the commission of offenses is the general form
of perpetration-through-another. To impose criminal liability for the per-
petration of an offense-through-another, it is necessary to prove aware-
ness of the instrumental use. Therefore, perpetration-through-another is
applicable only in mens rea offenses. In most cases, the person being in-
strumentally used by the perpetrator is considered an innocent agent, and
no criminal liability is imposed on him or her. We have already analyzed
perpetration-through-another in a context of mens rea offenses.
The person who is instrumentally used, however, may also be con-
sidered a semi-innocent agent who is criminally liable for negligence, al-
Table 4.2 . The Legal State Reflected in the Criminal Mental

State of the Other Person in Perpetration-through-Another
criminal men tal sta te

legal sta te of the o ther per son
of the o ther per son
mens rea accomplice

negligence semi-innocent agent
strict liability innocent agent
none innocent agent
though the perpetrator is at the same time criminally liable for a mens rea
offense. Negligence is the lowest level of mental element required for the
person instrumentally used to be considered a semi-innocent agent. In this
context, strict liability is too low a level of mental element for consider-
ation as a semi-innocent agent. If the person who is instrumentally used by
another is in a mental state of strict liability, that person is to be considered
an innocent agent, as if lacking any criminal mental state at all.
Thus, in perpetration of an offense-through-another, the other (instru-
mentally used) person may be in four possible mental states that reflect
matching legal consequences, as described schematically in table 4.2.
When the other person is aware of the delinquent enterprise and still
continues to participate although he is under no pressure, he becomes an
accomplice to the commission of the offense. Negligence reduces the per-
son’s legal state to that of a semi-innocent agent.36 But whether the mental
state of this person is that of strict liability or an absence of a criminal
mental state, he must be considered an innocent agent, so that no criminal
liability is imposed on him, and the full criminal liability for the rele
vant offense is imposed on the perpetrator who instrumentally used that
person. This is true for both humans and ai entities.
For this legal construct of perpetration-through-another, it is immate-
rial whether the ai system was instrumentally used, and whether it used
its strong ai capabilities to form a strict liability mental state. Thus, the
instrumental use of a weak ai system, of a strong ai system that formed
strict liability, or of a screwdriver has the same legal consequences. The
entity making the instrumental use (human, corporation, or ai system) is
criminally liable for the commission of the offense in full, and the instru-
mentally used entity (human, corporation, or ai system) is considered an
innocent agent.
For example, the human user of an unmanned vehicle based on an ad-
vanced ai system instrumentally uses the ai system to cross an intersection

where the traffic light is red. The system has the capability to be aware of
the traffic light’s color, but in practice it is not aware of the red light. The
system does not take all reasonable measures to examine the data, such as
analyzing the color of the traffic light facing in its direction. Crossing an
intersection on a red light is a strict liability offense. In this instance, the
human is criminally liable for that offense as perpetrator-through-another.
Because the ai system was instrumentally used by the perpetrator, it is
considered an innocent agent, and therefore criminal liability is not im-
posed on it.

Probable consequence liability addresses the commission of unplanned
offenses (different from a planned offense or additional to it). The ques-
tion in these cases is about the criminal liability of the other parties to
the unplanned offense committed by one party. Previously, we discussed
probable consequence liability for unplanned mens rea and negligence of-
fenses.37 The relevant question here is regarding the probable consequence
liability for unplanned strict liability offenses committed by an ai system.
Are the programmers, users, and other related persons criminally liable for
unplanned strict liability offense committed by an ai system?
For example, two people are committing a bank robbery. To escape
from the arena, they use an unmanned vehicle based on an advanced
ai system. During their escape, the vehicle exceeds the legal speed limit,
committing a strict liability offense. Analyzing the records of the vehicle
shows that it meets the strict liability requirements of the offense. The
bank robbers did not program the vehicle to commit the offense, and did
not order it to do so. The question concerns the robbers’ criminal liabil-
ity for the strict liability traffic offense committed by the vehicle, in ad-
dition to their criminal liability for robbery. If the offenders had ordered
the vehicle to drive at that speed and had instrumentally used it for that
purpose, this would have been perpetration-through-another, but that is
not the case here.
Probable consequence liability may be relevant here for the criminal
liability of the bank robbers. The mental condition for probable conse-
quence liability requires the unplanned offense to be “probable” from the
point of view of the party that did not commit it. In other words, it is nec-
essary for that party to have been able to foresee and reasonably predict the
commission of the offense. Some legal systems prefer to examine actual
and subjective foreseeability (the party has actually and subjectively fore-
seen the occurrence of the unplanned offense), whereas others prefer to
evaluate the ability to foresee based on an objective standard of reasonabil-
ity (the party has not actually foreseen the occurrence of the unplanned
offense, but any reasonable person under the same condition would have).
Actual foreseeability parallels subjective mens rea, whereas objective
foreseeability parallels objective negligence. A lower level of foreseeabil-
ity is not adequate for probable consequence liability. Thus, the question
is about the level of foreseeability required of the robbers. If they actually
foresaw the commission of that offense by the vehicle, then criminal li-
ability for the offense is imposed on them in addition to criminal liability
for robbery. If the robbers formed objective foreseeability, then criminal
liability for the additional offense is imposed only in legal systems in
which probable consequence liability can be satisfied through objective
foreseeability.
But if the mental state of the bank robbers in relation to the additional
offense is strict liability, this is not adequate for the imposition of criminal
liability through probable consequence liability. Although the additional
offense requires strict liability for imposition of criminal liability, this is
true only for the perpetrator of that offense, and not for imposing criminal
liability through probable consequence liability. Thus, if neither subjective
nor objective foreseeability regarding the commission of the offense can be
attributed to the robbers, then probable consequence liability is irrelevant.
In this type of cases, no criminal liability is imposed on the offenders, and
the criminal liability of the ai system for the strict liability offense does
not affect the offenders’ liability.
C l osin g the Opening Ex ample: Stric t L iable Killin g R obo t

We opened this chapter with an example of a killing robot, an unmanned
vehicle based on an ai system, subject to strict liability. Returning to the
example, we can now analyze it using the insights gained from this chap-
ter. The criminal liability of three main characters must be considered for
the homicide in this example: the ai system, the manufacturer (including
the programmers), and the user. To examine the criminal liability of the
ai system, first we must investigate its technological capabilities. If the
ai system has the technological capabilities needed to meet the require-
ments of strict liability offenses, it forms a machina sapiens criminalis in
this context.
If such capabilities are present, the criminal liability of the ai system
should be examined as if it were human. But if the ai system lacks the

technological capabilities needed to meet the requirements of strict liabil-
ity offenses, it does not form a machina sapiens criminalis in this context
and has no criminal liability for the offense. In the opening example, the
ai system met the factual element requirement of homicide, because its
conduct caused a chain of events that ended in the patient’s death.38
To meet the mental element requirement, we must prove strict liabil-
ity, which includes proving that the ai system is capable of consolidat-
ing awareness or negligence, and is capable of reasonability. To refute the
negligence presumption, it is necessary to prove that the ai system has not
actually formed awareness of relevant facts or negligence toward them,
and has taken all reasonable measures to prevent the commission of the
offense. The first condition may be proven based on the characteristics of
the system, using the definitions of awareness and negligence in criminal
law discussed earlier.39 If the system possesses such capability, and if such
capability has been activated, the system formed awareness or negligence.
The capability for reasonability is not different from the required par-
allel capability for negligence, previously discussed.40 The system is not
required to act reasonably, only to possess the capability of making rea-
sonable decisions. Strong ai systems, which can calculate the probabili-
ties of the occurrence of events based on current factual data (such as
weather forecast computers41) and choose options that meet certain crite-
ria, are supposed to have this type of capability. If no additional evidence
is brought before the court, this is sufficient to activate the negligence pre-
sumption and to impose criminal liability for the strict liability offense.
Refuting the negligence presumption begins with positive proof of not
having formed awareness or negligence. This may be proven based on the
records of the ai system. If proven, it is necessary to prove having taken all
reasonable measures for the prevention of the offense. Because the capabil-
ity for reasonability has already been proven, the conduct of the system is
examined accordingly. The reasonable options calculated by the system
and the actions it took based on these calculations must be examined to
determine whether they include all the reasonable options in this case.
The examination is carried out the same way as in the case of negligence,
at times using expert witnesses. If it is proven that no actual awareness or
negligence were formed regarding the factual data, and that all reasonable
measures have been taken by the system, the negligence presumption is
refuted. If not, criminal liability for the strict liability offense is imposed.
The criminal liability of the manufacturer, the programmer, and the
user is determined based on their role in causing the death. If they instru-
mentally used the ai system to kill the patient in the ambulance by design-
ing the system in such a way that it would collide with the ambulance
because it misunderstands the factual situation, they are perpetrators-
through-another of homicide. This is also the case if the ai system has
no advanced ai capabilities to meet the requirements of the offense. In
this case, they and the related parties, but not the ai system, are the only
ones criminally liable for the homicide. Their exact criminal liability for
the homicide, whether as murder or manslaughter, is determined by their
meeting the mental element requirements of these offenses.
If the manufacturer, the programmer, and the user foresaw the conse-
quences (the patient’s death) or any reasonable person in their condition
could have foreseen these consequences, and if the ai system is criminally
liable for the homicide, they may be criminally liable for the homicide
through probable consequence liability, as long as the offense was un-
planned. Generally, if they did not design the system to commit any of-
fense, but the unplanned homicide occurred, then the criminal liability of
the system is examined based on its capabilities and on meeting the rel-
evant homicide requirements, and the criminal liability of the human par-
ties is examined based on the standards of probable consequence liability.
Therefore, ai systems can be criminally liable for strict liability of-
fenses, and that does not necessarily reduce the criminal liability of other
relevant entities.
5
A pplica bili ty of G eneral Defen se s
to ai C riminal L ia bili ty
Can an ai system be insane? An infant? Intoxicated? Caught in situations

of self-defense, necessity, or duress? One of the most common uses of
ai technology is for guarding, as noted in chapter 1.1 For example, the
South Korean government uses ai robots as soldiers along its border with
North Korea, and since 2012 as prison guards.2 This technology faces par-
ticular threats to the software, aimed at weakening its ability to discharge
its guarding functions. This technology may also face dilemmas of the type
humans encounter in everyday life. If a prison guard apprehends an escap-
ing prisoner and the only way to stop him is by causing injury, is it right
for the robot to injure a human for that purpose?
Consider a case in which hackers manage to break into the main
computer controlling robots that function as prison guards. The hackers’
purpose is to help a friend escape from prison by incapacitating some
of the robots. The hackers plant a virus in the system, which immedi-
ately affects the robots, preventing them from forming the correct image
of the factual reality, and causes them to ignore some of the rules of the
system. Some of the prisoners use this situation to attempt an escape.
One prisoner captures a human guard and takes his uniform, trying to
escape by wearing it. This prisoner is caught by a human guard, how-
ever, and a struggle ensues. A robot witnessing the situation misinter-
prets it as a struggle between two prisoners, and approaches the two men
with the intention of intervening to stop the violence. When the human
guard orders the robot to keep away, the robot interprets this as a threat
to its mission, and grabbing the guard with its hydraulic arm, the robot
smashes him against the wall. The disguised prisoner escapes, and the
guard dies. The legal question is, who is to be held criminally liable for
the guard’s death?
The software controlling these robots is considered strong ai technol-
ogy, so the robots are capable of consolidating awareness and mens rea in
120
Applicabili ty of General Defenses . 121
the context of criminal law. Analysis of the records of these robots reveals
that the mens rea requirements of homicide are met in full. But this is not
the regular course of conduct for these robots, which are programmed not
to be violent unless it is absolutely necessary. Accurate analysis of this
case requires the application of the general defenses of criminal law to
ai technology. This is the question that this chapter seeks to answer: are
general defenses applicable to ai criminal liability?
5.1. The F unc tion of General Defenses in C ri minal L aw

General defenses in criminal law are complementary to the mental ele-
ment requirement; both address the offender’s fault in the commission
of the offense. The mental element requirement represents the positive
aspect of the fault (what should be present in the offender’s mind during
the commission of the offense), and the general defenses are the nega-
tive aspect of the fault (what should be missing from the offender’s mind
during the commission of the offense).3 For example, awareness is part
of the mental element requirement (mens rea), and insanity is a general
defense. Thus, in mens rea offenses, the offender must be aware and must
not be insane.
Therefore, the fault requirement in criminal law consists of the mental
element requirement and the general defenses. The general defenses were
developed in the ancient world to prevent injustice in certain types of
cases. For example, someone who killed another out of self-defense was
not criminally liable for the homicide, because the person lacked the re-
quired fault to cause death. An authentic factual mistake by the offender
regarding the intentional commission of an offense was considered as ne-
gating the required fault for imposing criminal liability.4 In the modern era,
the general defenses became wider and more conclusive, but the common
factor of all general defenses remained the same.
All general defenses in criminal law are part of the negative aspect of
the fault requirement, as they are intended to negate the offender’s fault.
The deep abstract question behind the general defenses is whether the
commission of the offense was not imposed on the offender in some way.
For example, when a person really acts in self-defense, the offense is con-
sidered to be imposed on her. To save her own life, which is considered a
legitimate purpose, the offender had no choice but to act in self-defense.
Naturally, she could have chosen to give up her life, but that is not con-
sidered to be a legitimate requirement because it goes against the natural
instinct of every living creature.
General defenses can be divided into two main types: exemptions and
justifications.5 Exemptions are general defenses related to the personal
characteristics of the offender (in personam), and justifications are related
to the characteristics of the factual event (in rem). The personal charac-
teristics of the offender may negate the fault for the commission of the
offense, regardless of the factual characteristics of the event or the exact
identity of the offense. In exemptions, the personal characteristics of the
offender are sufficient to prevent imposition of criminal liability for any
offense.
For example, a child under the age of legal maturity is not criminally
liable for any offense factually committed by him. The same is true for in-
dividuals who committed the offense at a time when they were considered
to be insane. The exact nature of the offense committed by the individual
is immaterial for the imposition of criminal liability. It may be relevant for
subsequent steps toward the treatment or rehabilitation of the offender, but
not for the imposition of criminal liability.
Justifications are impersonal general defenses, and as such they do not
depend on the identity of the offender, but only on the factual event that
took place. The personal characteristics of the individual are immaterial
for justifications. For example, one person is attacked by another person,
and her life is in danger. She pushes the attacker away, which is consid-
ered assault unless it is done with consent. In this case, the person doing
the pushing claims self-defense regardless of her identity, the identity of
the attacker, or any other of their personal attributes, because self-defense
has to do with only the factual event itself.
Because justifications are impersonal, they also have a prospective
value. Not only is the individual not criminally liable if she acted with
justification, but this is the way in which she should have been acting. Jus-
tifications define not only types of general defense, but also proper behav-
ior.6 Therefore, individuals should defend themselves in situations that
require self-defense, even if this may appear as an offense. This is not true
for exemptions. A child below the maturity age is not supposed to commit
offenses even if no criminal liability is imposed on him, and neither are
insane individuals.
The prospective behavioral value of justifications expresses the social
values of society. If we accept self-defense, for example, as legitimate jus-
tification, this means that we prefer people to protect themselves when
the authorities are unable to protect them. We prefer to reduce the monop-
oly of the state on power by legitimizing self-assistance rather than leave
people vulnerable and helpless. We do not coerce individuals to act in
their self-defense, but if they do, then we do not consider them criminally
liable for the offense committed through self-defense.
Both exemptions and justifications are general defenses. The term gen-
eral defenses relates to defenses that may be applied to any offense, and
not to a certain group of offenses. For example, infancy can be applied to
any offense as long as it was committed by an infant. By contrast, some
specific defenses can be applied only to specific offenses or types of of-
fenses. For example, in some countries in the case of statutory rape, a
defense is applicable if the age gap between the defendant and the victim
is less than three years. This defense is unique to statutory rape, and it is
irrelevant for any other offense. Exemptions and justifications are classi-
fied as general defenses.
As defense arguments, general defenses are argued positively by the
defense. If the defense chooses not to raise these arguments, they are not
discussed in court, even if participants at the trial understand that such an
argument may be relevant. It is not sufficient to state the general defense:
its elements must be proven by the defendant. In some legal systems, the
general defense is proven by raising a reasonable doubt about the presence
of the elements of the defense. In other legal systems, the general defense
must be proven by a preponderance of the evidence. Thus, the prosecution
has the opportunity to refute the general defense.
General defenses of the exemption type include infancy, loss of self-
control, insanity, intoxication, factual mistakes, legal mistakes, and sub-
stantive immunity. General defenses of the justification type include self-
defense (including defense of dwelling), necessity, duress, superior orders,
and de minimis defense. All these general defenses may invalidate the
offender’s fault. The question is whether these general defenses are appli-
cable to ai technology in the context of criminal law. We answer this ques-
tion now, discussing exemptions and justifications separately.
5.2. Exemption s
Exemptions are general defenses related to the personal characteristics
of the offender (in personam), as noted previously.7 The applicability of
exemptions to the criminal liability of ai entities raises the question of the
capability of these entities to form the personal characteristics required
for general defenses. For example, the question of whether an ai system
could be insane is rephrased to ask whether it has the mental capability of
forming the elements of insanity in criminal law. This question, mutatis
mutandis, is relevant to all exemptions, as we will discuss.
5.2.1. I nfancy
Could an ai system be considered an infant for the purposes of criminal
liability? Since ancient times, children under a certain biological age were
not considered criminally liable (doli incapax). The difference between
the various legal systems was in the exact age of maturity. For example,
Roman law set the age of maturity at seven years.8 This defense is deter-
mined by legislation9 and case law.10 It is clear that the cutoff age is biologi-
cal and not mental, primarily for evidentiary reasons, because biological
age is much easier to prove.11 It was presumed, however, that the biological
age matches the mental age.
If the child’s biological age exceeds the lower threshold but is under the
age of full maturity (for example, fourteen in the United States), the mental
age of the child is examined based on evidence (for example, expert testi-
mony).12 The conclusive examination is whether the child understands his
own behavior and its wrong character,13 and whether he understands that
he may be criminally liable for the offense, as if he were mature. But there
may still be some procedural changes in the criminal process compared to
the standard process (for example, juvenile court, the presence of parents,
or more lenient punishments).
The rationale behind this general defense is that children under a cer-
tain age, whether biological or mental, are presumed to be incapable of
forming the fault required for criminal liability.14 The child is not inca-
pable of mentally containing the fault and understanding the full social
and individual meanings of criminal liability, so that imposition of crimi-
nal liability would be irrelevant, unnecessary, and vicious. Consequently,
children are not criminally liable but rather are educated, rehabilitated,
and treated.15 The question, in our case, is whether this rationale is rele
vant only to humans but to other legal entities as well.
The general defense of infancy is not considered applicable to corpo-
rations. There are no “child corporations,” so the moment a corporation
is registered it exists, and criminal liability may be imposed on it legally;
therefore, the rationale of this general defense is irrelevant for corpora-
tions. A child lacks the mental capability to form the required fault be-
cause human consciousness is underdeveloped at this age. As the child be-
comes older, his or her mental capacity develops gradually until it reaches
full understanding of right and wrong. At this point, criminal liability
becomes relevant.
The mental capacity of corporations does not depend on their chrono-
logical “age,” that is, the date of registration, and it is considered to be
constant. Moreover, the mental capacity of a corporation derives from its
human officers, who are mature entities. Consequently, there is no legiti-
mate justification for the general defense of infancy to be applied to cor-

porations. The question of interest to us is whether ai systems resemble
humans or corporations in this context.
The answer differs for different types of ai systems, and we must dis-
tinguish between fixed ai systems and dynamically developing ones. Fixed
ai systems begin their activity with the capacities they are going to have
throughout their life cycle. These systems do not experience any change
in their capacities with the passage of time. Consequently, their capacity to
form mental requirements (for example, awareness, intent, or negligence)
must be examined at every point where an offense has been committed,
and the general defense of infancy does not apply.
But the starting and end points of dynamically developing ai systems
are different. Their capacities, including mental ones, develop over time
through machine learning or other techniques. If the system began its ac-
tivity without the mental capacities required for criminal liability, and at
some point it developed such capacities, then the time between the start-
ing point and the point when it possesses such capacities parallels that of
childhood. During this period, the system does not have the capabilities
required for the imposition of criminal liability.
But if the mental capacity of the ai system is already developed at the
time the offense is committed, the question arises whether the infancy
defense is still relevant because if the system possesses the required capa-
bilities, it is criminally liable regardless of the point in time when these
capabilities became available. The answer to this question is similar to the
rationale for the general defense for human offenders.16 We can examine
the mental capacity of each child at any age and determine individually
whether the child possesses the required capabilities, but this would be
very inefficient. Any preschooler who may have hit another one in day
care is immediately identified as not having the mental capacities required
for criminal liability.
The other way is to establish a legal presumption whereby children
under a certain age are not criminally liable. This is the situation with
ai systems. Massive production of strong ai systems (for example, dozens
of prison guards or military robots), having the exact same capabilities and
using the same learning techniques, makes it unnecessary to evaluate the
mental capacities of each individual robot. If we know empirically that a
given system possesses all required mental capabilities after 264 hours of
activity, and if such a robot commits an offense before having operated for
264 hours, the robot is presumed to be in its childhood and no criminal
liability is imposed.
If the prosecution insists that the individual system possesses the re-
quired mental capacities already in its “infancy,” or if the defense insists

that the system does not possess the required mental capacities despite
the fact that it has passed its infancy, the actual mental capacities of the
systems can be examined and evaluated specifically. This would not be
substantively different from the natural gaps between the biological and
mental ages of humans. When the argument is made that a seventeen-
year-old human offender is mentally underdeveloped, the court examines
the offender’s mental capacities and decides whether that person has the
required capacity to become criminally liable.
This rationale is relevant for both humans and dynamically developing
ai systems, but not for corporations. Therefore, it seems that the general
defense of infancy can be relevant for this type of ai systems under the
right circumstances.
5.2.2. L oss of Self-C ontrol

Could an ai system experience loss of self-control in the context of crimi-
nal liability? Loss of self-control is a general defense that has to do with
one’s inability to control one’s bodily movements. When the reason for this
inability is mental disease, it is considered insanity,17 and when it is the
effect of intoxicating substances, it is considered intoxication.18 But the
general defense of loss of self-control is more general than insanity and in-
toxication because it does not require a specific type of reason for the loss
of self-control. Whenever the offender’s bodily movements are not under
his or her full control, this defense may be used if its conditions are met.19
The general rationale of this defense is that uncontrollable bodily
movement does not reflect the offender’s will, and therefore it should not
be the basis for the imposition of criminal liability. Thus, the uncontrolla-
ble reflexes of an individual may be an expression of loss of self-control.20
For example, when a physician, during a routine medical checkup, taps
a patient’s knee, she triggers a reflex that causes the patient’s leg to move
forward. The leg may end up kicking the physician, which ordinarily may
have been considered assault, but because in this instance it is the result
of a reflex, the general defense of loss of self-control is applicable, and no
criminal liability is imposed.
But for the general defense to be applicable, the loss of self-control
must be total. In this example, if the patient tapped his own knee on pur-
pose for the reflex to be triggered, and as a result his leg kicked the physi-
cian, the general defense is not applicable because the assault reflects the
offender’s will. Consequently, two accumulative conditions must be met
for the loss of self-control defense to be applicable: (1) inability to control
the self-behavior, and (2) inability to control the conditions under which
that behavior occurred.
In our example, the patient did not control his self- behavior (the
reflex), but he controlled the conditions for its occurrence (tapping the
knee); therefore, the general defense of loss of self-control is not applicable
to him.
Many types of situations have been recognized as loss of self-control,
including automatism (acting without aware central control over the
body),21 convulsions, post-epileptic states,22 post-stroke states,23 organic
brain diseases, diseases of the central nervous system, hypoglycemia, hy-
perglycemia,24 somnambulism (sleepwalking),25 extreme sleep depriva-
tion,26 side effects of bodily27 or mental traumas,28 blackout situations,29
side effects of amnesia30 and brainwashing,31 and many more.32 The cause
for the loss of self-control is immaterial for meeting the first condition. As
long as the offender is indeed incapable of controlling his or her behavior,
the first condition is fulfilled.
The second condition refers to the cause for being in the first condi-
tion. If the cause was controlled by the offender, he is not considered to
have lost his self-control. Controlling the conditions for losing the self-
control is controlling the behavior. Therefore, when the offender controls
the conditions for moving in and out of control, he cannot be considered
to have lost his self-control. In Europe, the second condition is based on
the doctrine of actio libera in causa, which dictates that if the one controls
one’s entry into a situation that is out of control, the general defense of loss
of self-control is rejected.33
Accordingly, ai systems can experience loss of self-control in the con-
text of criminal law due to external and internal causes. For example, a
human pushes an ai robot onto another human. The robot being pushed
has no control over that movement. This is an example of an external
cause for loss of self-control. If the pushed robot makes nonconsensual
physical contact with the other person, this may be considered an assault.
The mental element required for assault is awareness, so if the robot is
aware of that physical contact, both factual and mental elements require-
ments of the assault are met.
If the ai robot were human, it would have probably claimed the gen-
eral defense of loss of self-control. Thus, although both mental and fac-
tual elements of the assault are fulfilled, no criminal liability is imposed
because commission of the offense was involuntary, or due to loss of self-
control. This general defense would prevent the imposition of criminal
liability on human offenders, and it should also prevent the imposition
of criminal liability on robots. If the robot has no ai capabilities for con-

solidating awareness, there is no need for this defense because the robot
would be functioning as a mere tool. The ai robot, however, is aware of
the assault.
The ability of the ai robot to meet the mental element requirement of
the offense makes it necessary to apply the general defense, which func-
tions identically with humans and robots. An example of internal cause
for loss of self-control is the case of an internal malfunction or technical
failure that causes uncontrolled movements by the robot. The robot may
be aware of the malfunction and still not be able to control it or correct it.
This is also the case for the general defense of loss of self-control. Whether
the cause for the loss of self-control is external or internal is relevant to the
applicability of this general defense.
If the robot controlled these causes, however, the defense is not ap-
plicable. For example, if the robot physically caused a person to push it
(the robot) onto another person (external cause), or if the robot caused the
malfunction knowing the probable consequences of this for its mecha-
nism (internal cause), then the second condition of the defense is not met
and the defense is not applicable. The situation is the same for humans.
It appears, therefore, that the general defense of loss of self-control is ap-
plicable to ai systems.
5.2.3. I nsanity
Can an ai system be considered insane for the purposes of criminal liabil-
ity? Insanity has been known to humanity since the fourth millennium
bc ,34 but at that time it was considered to be a punishment for religious
sins,35 and therefore there was no need to look for cures.36 Only since the
middle of the eighteenth century has insanity been explored as a mental
disease, and its legal aspects considered.37 In the nineteenth century, the
terms insanity and moral insanity described situations in which the indi-
vidual lost his or her moral orientation, or had a defective moral compass,
but was aware of common moral values.38
Insanity was diagnosed as such only in case of significant deviations
from common behavior, especially sexual behavior.39 Since the end of
the nineteenth century, it has been understood that insanity is mental
malfunctioning, and that at times it is not expressed in deviations from
common behavior. This approach lies at the foundation of the understand-
ings of insanity in criminal law and criminology.40 Mental diseases and
defects were categorized according to their symptoms and respective med-
ical treatments, and their effect on criminal liability was explored and re-
corded. But the different needs of psychiatry and criminal law produced
different definitions for insanity.
For example, the early English legal definition of an insane person
(“idiot”) was not being able to count to twenty. In psychiatry, such a
person is not considered insane.41 Criminal law needed a clear and con-
clusive definition of insanity, whereas psychiatry had no such needs. In
most modern legal systems, today’s legal definition of insanity is inspired
by two nineteenth-century English tests. One is the M’Naughten rules of
1843,42 and the second is the irresistible impulse test of 1840.43 The com-
bination of these two tests ensures that the general defense of insanity
complies with the structure of mens rea, as discussed previously.44
The legal definition of insanity has both cognitive and volitive aspects.
The cognitive aspect of insanity addresses the ability to understand the
criminality of the conduct, whereas the volitive aspect addresses the abil-
ity to control the will. Thus, if a mental disease or defect causes cognitive
malfunction (difficulty to understand the factual reality and the criminal-
ity of the conduct) or volitive malfunction (irresistible impulse), the condi-
tion is considered insanity from the legal point of view.45 This is the con-
clusive common test for insanity.46 It fits the structure of mens rea, which
also contains both cognitive and volitive aspects, and it is complementary
to the mens rea requirement.47
This definition of insanity is functional and not categorical. For an
individual to be considered insane, it is not necessary to have a mental ill-
ness that appears on a certain list of mental diseases. Any mental defect,
of any type, can be the basis for insanity as long as it causes cognitive or
volitive malfunctions. The malfunctions are examined functionally and
not necessarily medically, and they need not appear on a list of mental
diseases.48 Therefore, a person may be considered insane by criminal law
and perfectly sane from the point of view of psychiatry, as in the case of
a cognitive malfunction that is not categorized as a mental illness. The
opposite situation, whereby a person is sane for the purposes of criminal
law but insane psychiatrically, is also possible, as in the case of a mental
disease that does not cause any cognitive or volitive malfunction.
The insane person is presumed to be incapable of forming the fault re-
quired for criminal liability. The question is whether the general defense
of insanity is applicable to ai systems. The general defense of insanity re-
quires a mental, or inner, defect that causes cognitive or volitive malfunc-
tion. There is no need to demonstrate any specific type of mental disease,
and any mental defect is satisfactory. The question is, how can we know
about the existence of that “mental defect”? Because the mental defect is
examined functionally and not based on certain categories, the symptoms

of that mental defect are critical for its identification.
The inner defect causes cognitive or volitive malfunction, whether it
is classified as a mental disorder, or a chemical imbalance in the brain, or
an electric imbalance in the brain, and so on. The inner cause is exam-
ined based on its functional effect on the human mind. This is the legal
situation with humans, and the same holds true for ai systems. The more
complicated and advanced the ai system, the higher the probability of
inner defects mostly in the software, but also in the hardware. Some inner
defects do not cause the ai system to malfunction, but others do. If the
inner defect causes a cognitive or volitive malfunction of the ai system,
this matches the definition of insanity in criminal law.
Because strong ai systems are capable of forming all mens rea com
ponents, as we have discussed,49 and given that these components consist
of cognitive and volitive components owing to the mens rea structure, it
is quite likely that some inner defects can cause these capabilities to mal-
function. When an inner defect causes such a malfunction, this matches
the definition of insanity in criminal law. Partial insanity is applicable
when the cognitive or volitive malfunctions are not complete. Temporary
insanity is applicable when these malfunctions affect the offender (human
or ai system) for a certain period of time.50
It is possible to argue that this is not the typical character of the insane
person because it does not match the concept of insanity reflected in psy-
chiatry, culture, folklore, literature, and even the movies. Nevertheless,
it is insanity from the perspective of criminal law. First, the criminal law
definition of insanity differs from its definitions in psychiatry, culture, and
so on, and it is the definition that is used for humans. There is no reason
that a different definition should be used for ai systems. Second, criminal
law does not require a mental disease for human insanity, so why should
we require it for ai systems?
The definition of insanity used by criminal law may seem too techni-
cal, but if the offender meets its requirements, it is applied. If both human
and ai offenders are capable of meeting the insanity requirement in crimi-
nal law, there is no legitimate reason for making the general defense of
insanity applicable for one but not the other. Consequently, it appears that
the general defense of insanity is applicable to ai systems.
5.2.4. I ntoxication
Can an ai system be considered intoxicated for the purposes of criminal li-
ability? The effects of intoxicating substances have been known to human-
ity since prehistory. In the early law of ancient times, the term intoxication
referred to drunkenness as a result of ingesting alcohol. Later, when the
intoxicating effects of other materials became known, the term was ex-
panded.51 Until the beginning of the nineteenth century, intoxication was
not accepted as a general defense. The Archbishop of Canterbury wrote
in the seventh century that imposing criminal liability on a drunk person
who committed homicide was justified for two reasons: first, the drunken-
ness itself, and second, the homicide of a Christian person.52
Drunkenness was conceptualized as a religious and moral sin, and
therefore it was considered immoral to let offenders be exempt from crim-
inal liability for being drunk.53 Only in the nineteenth century did the
courts undertake a serious legal discussion of intoxication, made possible
by legal and scientific developments that have produced the understand-
ing that an intoxicated person is not necessarily mentally competent for
purposes of criminal liability (non compos mentis).
From the beginning of the legal evaluation of intoxication in the nine-
teenth century, the courts distinguished between cases of voluntary and
involuntary intoxication.54 Voluntary intoxication was considered to be
the offender’s fault, so it could not be the basis for exemption from crimi-
nal liability. Nevertheless, voluntary intoxication could be considered as a
relevant circumstance for the imposition of a more lenient punishment.55
Moreover, voluntary intoxication could be used to refute premeditation in
first-degree murder cases.56 Courts have distinguished between cases based
on the reasons for becoming intoxicated. Voluntary intoxication grounded
in the will to commit an offense was considered different from voluntary
intoxication undertaken for no criminal reason.57
By contrast, involuntary intoxication was recognized and accepted as
an exemption from criminal liability.58 Involuntary intoxication is a situa-
tion imposed on the individual, so it is not just and fair to impose criminal
liability in such situations. Thus, the general defense of intoxication has
two main functions. When intoxication is involuntary, the general defense
prevents imposition of criminal liability. When it is voluntary but not in-
tended for the commission of an offense, intoxication is a consideration
for the imposition of a more lenient punishment.
The modern understanding of intoxication includes any mental effect
caused by an external substance (for example, chemicals). The required
mental effect matches the structure of mens rea, discussed earlier. Conse-
quently, the effect may be cognitive or volitive.59 The intoxicating effect
may relate to the offender’s perception, understanding of factual reality,
or awareness (cognitive effect); or it may affect the offender’s will, includ-
ing the creation of an irresistible impulse (volitive effect). Intoxication is

caused by an external substance, but there is no closed list of substances
that may be illegal (heroin, cocaine, and so on) or perfectly legal (alcohol,
sugar, pure water, and so on).
The effect of the external substance on the individual is subjective,
and various people may be affected differently by the same substances in
the same quantities. Sugar may produce hyperglycemia in one person, but
another person is barely affected by it. Pure water may result in an imbal-
ance of the electrolytes in one person, but another is barely affected by it.
Cases of addiction raised the question whether the absence of the external
substance may also be considered a cause for intoxication. For example,
when a person addicted to narcotics is in withdrawal, he or she may expe-
rience cognitive and volitive malfunctions as a result of the drug’s absence.
Consequently, the cognitive and volitive effects of narcotics addiction
are considered intoxication by criminal law.60 To determine the question of
voluntary or involuntary intoxication (in case the addicted person wanted
to begin the process of weaning), it is the cause for the addiction, not for
the weaning, that is examined to be voluntary or involuntary.61 Thus, in-
toxication is examined through a functional evaluation of its cognitive and
volitive effects on the individual, regardless of the exact identity of the ex-
ternal substance that is the cause of these effects. The question is whether
the general defense of intoxication is applicable to ai systems.
As already noted, the general defense of intoxication requires an exter-
nal substance (for example, the presence or absence of a certain chemical)
that has cognitive or volitive effects on the inner process of conscious-
ness. For example, the manufacturer of ai robots wanted to reduce produc-
tion costs, and used inexpensive materials. After a few months of opera-
tion, a process of corrosion began in some of the initial components of the
robots, and as a result, transfer of information was deficient and affected
the awareness process. Technically, this is similar to the effect of alcohol
on human neurons.
In another example, a military ai robot was activated in a civilian area
after a real or simulated chemical attack. As a result of exposure to the
poisonous gas, parts of the robot’s hardware were damaged, and the robot
began to malfunction, became unable to properly identify people, and
started to attack innocent civilians. Subsequent analysis of the robot’s re-
cords showed that exposure to the gas was the only reason for the attacks.
If the robot had been human, the court would have accepted the general
defense of intoxication and acquitted the defendant. Should not the same
procedure also apply to ai systems?
If there is no difference between the effects of external substances on

humans and ai systems from a functional point of view, then there is no
justification for applying the general defense of intoxication to one and
not the other. Strong ai systems can possess both cognitive and volitive
inner processes, as noted previously.62 These processes can be affected by
various factors, and when they are affected by external substances, as in
the examples just presented, the requirements of intoxication as a general
defense are met. Thus, if exposure to certain substances affects the cogni-
tive and volitive processes of an ai system in a way that causes it to commit
an offense, there is no reason that intoxication as a general defense should
not be applicable.
It may be true that ai systems cannot become drunk on alcohol or have
delusions following the ingestion of drugs, but these effects are not the
only possible ones related to intoxication. If a human soldier attacks his
friends as a result of exposure to a chemical attack, his argument for in-
toxication is accepted. If this exposure has the same substantive and func-
tional effect on both humans and ai systems, there is no legitimate reason
for making the general defense of intoxication applicable to one type of
offender but not another. Consequently, it appears that the general defense
of intoxication can be applicable to ai systems.
5.2.5. F actual M istake

We are such stuff
As dreams are made on, and our little life
Is rounded with a sleep.
— William Shakespeare, The Tempest, 1611
Can an ai system be considered factually mistaken for the purposes of

criminal liability? The general defense of factual mistake provides a re-
vised perspective of the cognitive aspect of consciousness. In our discus-
sion of awareness, the assumption was that there is a factual reality of
which the individual may or may not be aware.63 This reality was consid-
ered constant, objective, and external to the individual. But the only way
an individual can know about the factual reality is through the process
of awareness, that is, perception by senses and understanding of the fac-
tual data.64
Only when our brain tells us that this is the factual reality do we be-
lieve it; we have no other way to evaluate reality’s existence. But sights,
sounds, smells, or physical pressure may be simulated, and the human
brain may be stimulated to feel them, although they may not necessarily
exist. The ultimate example is dreaming. Often, we dream without know-

ing that we are dreaming, and we cannot distinguish between dream and
factual reality while we are dreaming. In our dreams we see sights, hear
sounds, smell, feel, talk, or run as if we were facing factual reality.
What makes the dream different from factual reality in the eyes of
the dreamer? For most people, the dream is a dream simply because the
dreamer eventually exits it by waking up. But what would happen if, on
the contrary, we were to wake up in a dream? This option is not accepted
by most people because our intuition tells us that if we open our eyes, if
we are in bed, and if it is morning, the dream portion of this day has ended
and the waking portion begins. But when we are in the midst of the dream,
we have no reliable way of distinguishing dreams from what we call “fac-
tual reality” (within quotation marks, because we do not have a reliable
way of verifying it).
We “see” things with our brain, not with our eyes. Our eyes are merely
light sensors. The light is converted into electrical currents and chemi-
cal processes that carry messages between neurons and form the “sight”
by stimulating the brain in the right spots. But it is possible to imitate
this stimulation. Electrodes can be attached to the brain to stimulate it in
the appropriate regions and make it “see,” “hear,” or “feel.” The stimu-
lated brain becomes aware of these images, sounds, and so on, and people
would consider them factual reality.
If we begin to doubt our awareness of the factual reality, we must also
consider the problem of perspective. Even if we assume that we experi-
ence the factual reality (without quotation marks), we can experience it
only through our subjective perspective, which is not necessarily the only
perspective of that reality. For example, if we cut out a triangular shape
from a piece of cardboard, we may perceive a triangle if we view it facing
its flat surface; but if we rotate it by 90 degrees, we perceive it as a straight
line. The issue of perspective can be crucial if we associate different inter-
pretations with the different perspectives.
For example, two people see a man holding a long knife and hear him
telling a woman that he is about to kill her. One person understand this
as a serious threat to the woman’s life, and calls police or attempts to in-
tervene and save the woman; the other person understands this as part of
a show (for example, a street performance) that requires no intervention.
The deep question in criminal law in this context is, what should be the
factual basis of criminality: the factual reality as it actually occurred, or
what the offender believed to be the factual reality, although it may not
have actually taken place?
For example, a man charged with rape admits that he and the com-
plainant had sexual intercourse, but proves that he had really believed
that it was carried out consensually; and the prosecution proves that the
complainant did not consent. The court believes both, and both are telling
the truth. Should the court acquit or convict the defendant? Since the sev-
enteenth century, modern criminal law has preferred the subjective per-
spective of the individual (the defendant) about the factual reality to the
factual reality itself.65 The general concept is that the individual cannot be
criminally liable except for “facts” that he or she believed to be knowing,
regardless of whether they actually occurred in factual reality.66
In most legal systems, the limitations imposed on this concept are evi-
dentiary, so that the defendant’s argument would be considered true and
authentic, if proven by the defendant. When the argument is considered
true and authentic, this becomes the basis for the imposition of crimi-
nal liability. In our example, the defendant is acquitted because he truly
believed the intercourse was consensual. If the defendant’s perspective
negates the mental element requirement, the factual mistake works as a
general defense that leads to the defendant’s acquittal.67
The general defense of factual mistake is applicable not only in mens
rea offenses, but also in negligence and strict liability offenses. The differ-
ence between them is in the type of mistake required. In mens rea offenses,
any authentic mistake negates awareness of the factual reality and is con-
sidered adequate for that general defense. In negligence offenses, the mis-
take is also required to be reasonable for the defendant to be considered
to have acted reasonably.68 In strict liability offenses, the mistake must be
inevitable despite the defendant having taken all reasonable measures to
prevent it.69 The question is whether the general defense of factual mistake
is applicable to ai systems.
Both humans and ai systems can experience difficulties, errors, and
malfunctions in the process of awareness of the factual reality. These dif-
ficulties may occur both in the process of absorbing the factual data by the
senses and in the process of creating the relevant general image about the
data. In most cases, such malfunctions result in the formation of an inner
factual image that differs from the factual reality as the court understands
it. This is a factual mistake concerning the factual reality. Factual mistakes
are part of our everyday life, and they are a common basis for our behavior.
In some cases, factual mistakes by both humans and ai systems can
lead to the commission of an offense. This means that according to factual
reality the action is considered an offense, but not so according to the sub-
jective inner factual image of the individual, which happens to involve a
factual mistake. For example, a human soldier mistakenly identifies his

friend as an enemy and shoots him. For reasons unknown, the soldier
who was shot wore the enemy’s uniform, spoke the enemy’s language, and
looked as if he intended to attack the soldier who shot him. Although he
was asked to identify himself, he did not comply with the request. In this
case, the mistake is authentic, reasonable, and inevitable.
If the human soldier claims a factual mistake, he will probably be
acquitted (if indicted at all). If the soldier is not a human but a strong
ai robot, should not the criminal law treat the robot soldier the same way
that it treats the human soldier? The errors committed by the human and
robot soldiers are substantively and functionally identical. The factual
mistakes of both humans and ai entities have the same substantive and
functional effects on cognition and on the perception of factual reality. As
a result, there is no reason for preventing the use of the factual mistake as
general defense for ai entities exactly as it is applied to humans.
Computers do make mistakes. The probability of a mistake in math-
ematical calculations by a computer may be low, but if the computer ab-
sorbs mistaken factual data, the results of the calculations may be wrong.
Calculations regarding required, possible, and impossible courses of action
are affected by the data being absorbed, exactly as in the case of humans.70
If factual mistakes have identical substantive and functional effects on
humans and ai systems, there is no legitimate reason for making the gen-
eral defense of factual mistake applicable to one type of offender and not
to the other. Consequently, it appears that the general defense of factual
mistake can be applied to ai systems.
5.2.6. L egal M istake

Could an ai system be considered legally mistaken for the purpose of crim-
inal liability? A legal mistake is a situation in which the offender either
misinterprets the law or is ignorant of it. The general idea behind this de-
fense is that a person who does not know about a certain prohibition, and
consequently commits an offense, does not consolidate the required fault
for the imposition of criminal liability. Nevertheless, the mental element
of the offenses does not include knowing about the prohibition.
For example, the offense of rape requires a mental element of mens rea,
which includes awareness of the commission of sexual intercourse with a
woman and of the absence of consent. That offense does not require that
the rapist know that rape is prohibited as a criminal offense. The mental
element requirement of rape is satisfied by awareness of the factual ele-
ment components, regardless of the rapist’s knowledge that rape is pro-
hibited. This is the legal situation for most offenses owing to prospective
considerations of everyday life in society.71
If offenders were required to know about the prohibition as a condi-
tion for the imposition of criminal liability, they would be encouraged not
to learn the law, and as long as they were ignorant of the law, they would
enjoy immunity from criminal liability. If no such condition is required,
the public is encouraged to inquire about the law, to know it, and to obey
it. These prospective considerations do not fully match the fault require-
ment in criminal law. As a result, criminal law seeks a balance between
justice, the fault requirement, and prospective considerations concerned
with everyday life in society.
Initially, considerations were entirely prospective. Roman law stated
that ignorance of the law does not excuse the commission of offenses (ig-
norantia juris non excusat),72 and until the nineteenth century, this was
the general approach followed by most legal systems.73 In the nineteenth
century, when the culpability requirement in criminal law underwent a
dramatic development, it became necessary to establish a balance, and the
legal mistake was required to be made in good faith (bona fide)74 and to
reflect the highest state of mental element, that is, strict liability. Accord-
ing to this standard, the legal mistake is an inevitable one, despite the fact
that all reasonable measures have been taken to prevent it.75
This high standard for legal mistakes is required for all types of of-
fenses, regardless of their mental element requirement. Therefore, the gen-
eral standard for legal mistakes is higher than that for factual mistakes.
In this context, the main question in courts is whether the offender has
indeed taken all reasonable measures to prevent the legal mistake, includ-
ing reasonable reliance on statutes, judicial decisions,76 official interpreta-
tions of the law (including pre-rulings),77 and the advice of private coun-
sel.78 The question for us is whether the general defense of legal mistake is
applicable to ai systems.
Technically, if the relevant entity, whether a human or an ai system,
is capable of fulfilling the mental element requirement of strict liability
offenses, the entity also is capable of claiming legal mistake as a general
defense. Because strong ai systems have the ability to meet the mental el-
ement requirement of strict liability offenses, as discussed earlier,79 they
also have the capabilities needed to make the general defense of legal mis-
take applicable to them. The absence of the legal knowledge of a given
issue can be proven by examining the records attesting to the knowledge
of the ai entity, thereby meeting the good faith requirement as well.
The fundamental meaning of the applicability of the legal mistake
defense to ai systems is that the system was not restricted by any formal
legal restriction, and it acted accordingly. If the ai system contains a soft-
ware mechanism that searches for such restrictions, and although the
mechanism has been activated no such legal restriction has been found,
then the general defense applies. Note, however, that the system’s ex-
emption from criminal liability does not function as an exemption from
criminal liability for the programmers or users of the system. If these
persons could have restricted the activities of the system to strictly legal
ones but failed to do so, they may be criminally liable for the offense
based on perpetration-through-another or on probable consequence, as
we have discussed.80
For example, an ai system absorbs factual data about certain people,
and it is required to analyze their personalities based on the data and pub-
lish the results within certain guidelines. In one instance, the publication
is considered criminal libel. If the records of the system show that it has
not been constrained by any restriction regarding libelous publications
and it either did not have the mechanism to search for such restrictions
or it did have the mechanism but found no such restrictions, then the
system is not criminally liable for libel. The manufacturer, programmers,
and users may be criminally liable for the libel, however, as perpetrators-
through-another or through probable consequence liability.
An ai system may have broad knowledge about many types of issues,
but it may not contain legal knowledge on every legal matter. The system
may be searching for legal restrictions if it is designed to do that, but may
not necessarily find any; this then becomes a case for the general defense
of legal mistake. If legal mistakes have the same substantive and functional
effects on both humans and ai systems, there is no legitimate reason to
make the general defense of legal mistake applicable to one type of of-
fender but not to the other. Consequently, it appears that the general de-
fense of legal mistake is applicable to ai systems.
5.2.7. Substantive I mmunity

Could an ai system have substantive immunity in the context of criminal
law? Certain types of people enjoy substantive immunity from criminal
liability by virtue of their office (ex officio). The immunity is granted to
these people ex ante so that they are not troubled with issues of criminal
law related to their office. Society grants these immunities because the of-
fices held by such people are regarded to be much more important than
the probable criminal offenses they may commit in the course of discharg-
ing their duties. This immunity is not absolute, and it relates to offenses
committed by these people as part of discharging their official duty and

for its sake.
For example, a firefighter on a mission to save a woman’s life is hoisted
by a crane and breaks the window to enter her apartment. This act ap-
parently meets the factual and mental element requirements of several
offenses (for example, break-in, entering, and property damage), but to
enable the firefighter to discharge his lifesaving duty, he is granted sub-
stantive immunity from criminal liability for these offenses. If, however,
the firefighter had entered the apartment and raped the woman, the im-
munity would have been immaterial because rape is not part of his duties.
This type of immunity is substantive as it nullifies the criminal liabil-
ity of the person to whom it is extended, but it does not prevent indictment
or other criminal proceedings. The law explicitly lists the types of people
to whom substantive immunity is extended (firefighters, police officers,
soldiers, and so on). The general defense applies to these people only if
they have committed the offense as part of their official duty, in order to
discharge that duty, and in good faith (bona fide), that is, not exploiting
the immunity to deliberately commit other criminal offenses. The ques-
tion is whether the general defense of substantive immunity is applicable
to ai systems.
The firefighter in our example was human, and if indicted for causing
property damage would claim substantive immunity. The court would
most likely accept the claim and acquit him immediately. Let us assume,
however, that the fire posed too heavy a risk for a human firefighter, and
an ai robot was dispatched to save the woman’s life. The robot is equipped
with a strong ai system and capable of making the same decision that
a human firefighter would make under the circumstances. Is there any
reason such a robot should not be granted similar immunity to that en-
joyed by the human firefighter?
If all the conditions for granting this general defense are met, there is
no reason for using different standards toward humans and ai systems. ai
robots are already being used for official duties (for example, as guards),81
and inevitably at times they must commit offenses in the course of dis-
charging their duties. For example, prison guards may physically assault
escaping prisoners to prevent them from escaping. If these situations have
the same substantive and functional effects on both humans and ai sys-
tems, then there is no legitimate reason for make the general defense of
substantive immunity applicable to one type of offender but not to the
other. Consequently, it appears that the general defense of substantive im-
munity is applicable to ai systems.
5.3. Jus tifica tion s

Justifications are general defenses having to do with the characteristics of
the factual event (in rem), as previously noted.82 The applicability of justi-
fications to the criminal liability of ai systems raises the question of the ca-
pability of these systems to be involved in such situations as self-defense,
necessity, or duress. Is it legitimate to allow a robot to defend itself from
attack? And if the attack is conducted by humans, is it legitimate to allow
a robot to attack humans to ensure the robot’s safety?
In general, because justifications are in rem general defenses, the per-
sonal characteristics of the individual (human or ai system) should be con-
sidered immaterial. But these general defenses were designed for humans,
taking into account human weaknesses and allowing for them. For ex-
ample, self-defense was designed to protect the human instinct to protect
one’s life. Is this instinct relevant to ai systems, which are machines? If
not, why would self-defense apply to machines? We will now explore the
applicability of justifications as general defenses for ai systems.
5.3.1. Self-Defense
Self-defense is one of the most ancient defenses in human culture. Its
basic function is to mitigate the society’s absolute monopoly on power
according to which only society (the state) has the authority to use force
against individuals.83 When one individual has a dispute with another, he
or she is not allowed to use force, but must apply to the state (through the
courts, police, and so on) for the state to solve the problem and use force
if necessary. According to this concept, no power is left in the hands of
individuals.
For this concept to be effective, however, state representatives must be
present at all times in all places. If one individual is attacked by another
in a dark alley, she cannot retaliate, but must wait for the state represen-
tatives, who may be unavailable at that time. To enable people to protect
themselves from attackers in situations like this, society must partially re-
treat from this concept. One such retreat is the acceptance of self-defense
as a general defense in criminal law. Self-defense allows individuals to
protect some values using force, skirting the monopoly on power enjoyed
by the state.
Being in a situation that warrants self-defense is considered to nul-
lify the individual’s fault required for the imposition of criminal liability.
This concept has been accepted by legal systems worldwide since ancient
times, and eventually the defense has become wider and more accurate.84
Its modern form allows individuals to repel attacks on legitimate interests.

There are several conditions for this general defense to be applicable:
1. The protected interest must be legitimate. Legitimate interests

are the life, freedom, body, and property85 of the individual or
of other individuals.86 No previous introduction between the
person threatened and the defender is required,87 so that self-
defense is not only of the “self.”
2. The protected interest must be attacked illegitimately.88 When a
police officer, carrying a legal warrant, attacks an individual to
make an arrest, the attack is considered legitimate.89
3. The protected interest must be in clear and imminent danger.90
4. The act of self-defense must be carried out to repel the attack,
and it must be proportional to it,91 and immediate.92
5. The defender must not be in a situation to control the attack or
the conditions for its occurrence (actio libera in causa).93
If all these conditions are met, the individual is considered to be acting in

self-defense, and therefore no criminal liability is imposed on him or her
for the commission of the offense. But not every time an attack is repelled
can it be considered self-defense; only when the act meets the listed con-
ditions in full. The question is whether the general defense of self-defense
is applicable to ai systems. The answer depends on the capabilities of the
ai systems to meet the conditions just outlined, in the relevant situations.
At the heart of self-defense is the legitimate protected interest. When
the protected interest is that of another person (a human interest pro-
tected by the ai system using self-defense), there seems to be no difficulty.
Indeed, this is the broad basis for the legal activity of guard robots (of
humans, prisoners, borders, dwellings, and so on). If these robots have the
authority to repel attacks on human interests, self-defense is a legal justi-
fication. But the question is whether the self-interest of an ai system may
also be legally protected by self-defense. Indeed, there are two questions
here: one is moral, the other legal.
Morally, the question is whether we accept the idea of a robot pro-
tecting itself and having derivative rights, some of them constitutional.94
The moral question has nothing to do with the legal one. Since the 1950s,
the human approach to this moral question has generally been positive.
Recall Asimov’s third law, which states, “A robot must protect its own
existence, as long as such protection does not conflict with the First or
Second Laws.”95 According to this law, under the right circumstances the
robot is not only authorized to protect itself, but it must do so. As already
noted, however, this is not part of the legal question, and as far as the ap-
plicability of self-defense to robots is concerned, only the legal question
is relevant.
The legal question is much simpler, and it asks whether an ai system
has the ability to protect its own “life,” freedom, body, and property. To
protect these legitimate interests, the ai system must first possess them.
Only if an ai system possesses property can it protect it; otherwise it can
protect only the property of others. The same is true concerning life, free-
dom, and body. Currently, criminal law protects human life, freedom,
and body. The question of whether a robot has analogous life, freedom,
and body is one of legal interpretation, in addition to the moral questions
involved.
Criminal law recognized decades ago that the corporation, which is not
a human entity, possesses life, freedom, body, and property.96 We return
to this analogy in the next chapter when we discuss the punishments of
ai systems, including capital punishment and imprisonment. But if the
legal question concerning corporations, which are abstract creatures, has
been decided affirmatively, it would be unreasonable to decide otherwise
in the case of ai systems, which physically simulate these human attri-
butes much better than do abstract corporations.
For example, in a prisoner escape, a prison-guard robot is attacked by
the escaping prisoners, who intend to tie it up in order to incapacitate it
and prevent it from interfering with their escape. Is it legitimate for the
robot to defend its mission and protect its freedom? If the prisoners intend
to break its arms, is it legitimate for the robot to defend its mission and
protect its “body”? If the answers are analogous to those supplied for cor-
porations, they are affirmative. In any case, even if the idea of an ai robot’s
life, body, and freedom is not acceptable, despite the analogy with corpo-
rations and despite the positive moral attitude, ai systems still meet the
first condition of self-defense by protecting the life, body, freedom, and
property of humans.
The other two legal conditions of self-defense do not appear to differ
for humans and ai systems. The condition requiring an illegitimate attack
depends on the attacker, not on the identity of defender, whether human or
not. The condition of clear and imminent danger to the protected interest
does not depend on the defender either. The nature of the danger is deter-
mined by the attacker, not by the defender, and therefore this condition is
met not through the behavior of the defender, but through the analysis of
the attack, independent of the defender’s identity.
The condition of repelling the attack by a proportional and immediate
action raises another question: is it legitimate for a robot to attack a human,

even if the attack is in order to repel, by a proportional and immediate re-
action, a human attack? This question raises two subsidiary questions. The
first involves the legitimacy of preferring the defender’s rights over those
of the attacker, regardless of their identity as humans, corporations, or
ai systems. The second concerns restriction on ai systems as defenders
that differ from the restrictions imposed on humans and corporations.
Both subsidiary questions have legal and moral aspects.
The first subsidiary question is generally answered with reference to
risk taking by the attacker. Society creates peaceful social mechanisms
for dispute resolutions, such as legal proceedings, arbitration, mediation,
and so on. The attacker chooses not to use these peaceful mechanisms, but
rather chooses to resort to illegal violence. When acting in this manner and
bypassing the legal mechanisms of dispute resolution, the attacker takes
the risk of triggering a response against the illegal attack. In this case, crim-
inal law prefers the innocent reaction over the illegal action.97 This answer
recognizes no difference between various types of attackers and defenders.
There may be several moral answers to the second subsidiary question,
but none is relevant to the legal answer. Therefore, despite Asimov’s first
law, noted earlier,98 which prohibits robots from harming humans, robots
are actually used in several ways that harm humans. Using robots in mili-
tary and police functions (as soldiers, guards, armed drones, and so on)99
inherently implies the possibility of causing harm to humans. As long as
this use has not been explicitly prohibited by criminal law, it is legitimate.
It appears that legal systems worldwide have not only made their legal
choice in this matter, but have made their moral choice as well, rejecting
Asimov’s first law, which has been regarded as too panicky.
Consequently, criminal law sees no legal problem with the possibil-
ity that a robot may protect legitimate interests under illegal attack by
humans. This question would not have been raised at all if the attacker
were a robot, in which case the answer would be that the attack must be
repelled with proportional and immediate action, whether the defender is
human or not. The reaction is evaluated identically, regardless of whether
the defender is human. Repelling means that the act is a reaction to the
attack, which is the cause for the reaction. Immediate means that the reac-
tion is subsequent to the attack. The reaction is not illegal revenge, but an
act intended to neutralize the threat.
Proportional means that the defender does not use excessive force to
neutralize the threat. The defender must evaluate the threat and the means
that may be used to neutralize it, and then choose the means that are not
excessive in relation to the specific threat under the given circumstances.

Proportionality resembles reasonability in many ways, and some under-
stand it as being part of reasonability. Therefore, to meet this condition, the
ai system must have the capability of reasonableness, which is required for
negligence as well, as discussed previously in that context.100
Finally, it is required that the defender have no control over the attack
or the conditions for its occurrence. This requirement is intended to
impose criminal liability on persons who brought the attack upon them-
selves and then attempt to save themselves from criminal liability by using
the self-defense argument. The court is likely to seek out the deep reasons
for the attack. If the defender was innocent in this context, the defense is
applicable. This requirement is not applied differently to humans, corpo-
rations, or ai entities. As long as the ai entity was not part of such a plot,
the general defense of self-defense is applicable to it. The proof may be
based on the ai system’s records. Consequently, it appears that the general
defense of self-defense is applicable to ai systems.
5.3.2. N ecessity
Can an ai system be considered to be acting out of necessity in the criminal
law context? Necessity is a justification of the same type as self-defense.
Both are designed to mitigate the society’s absolute monopoly on power,
discussed earlier.101 The principal difference between self-defense and ne-
cessity is in the identity of the object of the response. In the case of self-
defense, the defender’s reaction is against the attacker, whereas in the case
of necessity it is against an innocent object (innocent person, property,
and so on). The innocent object is not necessarily related to the cause of
the reaction.
For example, two people are sailing on the high seas. The boat hits an
iceberg and sinks. Both people survive on an improvised raft, but have
no water or food. After a few days, one of them eats the other in order to
survive.102 The victim did not attack the eater and was not to blame for the
crash; she was completely innocent. The attacker was also innocent, but
he knew that if he did not eat the other person, he would definitely die.
If the eater is indicted for murder, he can claim necessity (self-defense is
not relevant in this case because the victim did not threaten the attacker).
The traditional approach toward necessity is that under the right cir-
cumstances, it can justify the commission of offenses (quod necessitas non
habet legem).103 The traditional justification of the defense is the under-
standing in criminal law of the weakness of human nature. The individual
who acts under necessity is considered to be choosing the lesser of two
evils from his or her own point of view.104 In our example, if the attacker
chooses not to eat the other person, they both perish; if he eats the other
person, only one of them dies. Both situations are evil, but the lesser evil
of the two is the one in which one person survives. The victim of neces-
sity is not blamed for anything but is innocent, and the act is considered
to be justified.105
Given that the act performed out of necessity is the individual’s, car-
ried out when the authorities are not available, the general defense of ne-
cessity partially reduces the applicability of the general concept whereby
only society (the state) has a monopoly on power, as we have discussed.
Being in a situation that requires an act of necessity is considered to nul-
lify the individual’s fault required for the imposition of criminal liabil-
ity. This concept has been accepted by legal systems worldwide since
ancient times, and eventually the defense became broader and more ac-
curate, so that its modern form allows individuals to protect legitimate
interests.106

are the life, freedom, body, and property of the individual or
of other individuals. No previous introduction between the
person threatened and the defender is required.107
3. The act of necessity must be directed against an external and
innocent interest.109
4. The act of necessity must be carried out to neutralize the danger,
and it must be proportional to it,110 and immediate.111
5. The defender must not be in a situation to control the danger or
the conditions for its occurrence (actio libera in causa).
If all these conditions are met, the individual is considered to be acting

under necessity, and therefore no criminal liability is imposed on him or
her for the commission of the offense. But not every danger neutralized
by causing harm to an innocent interest can be considered necessity; only
when the act meets the listed conditions in full. The question is whether
the general defense of necessity is applicable to ai systems. The answer
to this question depends on the capabilities of the ai systems to meet the
conditions just outlined, in the relevant situations.
Four of the listed conditions — numbers 1, 2, 4, and 5 — are identical
with those of self-defense, mutatis mutandis; the difference is that instead
of an attack on a legitimate interest, we are facing danger to that same in-
terest. The main difference between self-defense and necessity lies in one
of the conditions. Whereas in self-defense the act is directed toward the
attacker, in the case of necessity the act is directed toward an external or
innocent interest. In the case of necessity, the defender must choose be-
tween the lesser of two evils, one of which harms an external interest, who
may be an innocent person having nothing to do with the danger.
The question regarding ai systems in this context is whether they are
capable of choosing the lesser of two evils. For example, an ai -based lo-
comotive transporting twenty passengers arrives at a junction of two rail-
roads. A child is playing on one of the rails, but the second rail leads
to a cliff. If the system chooses the first rail, the child will certainly die
but the twenty passengers will survive. If it chooses the second rail, the
child will survive, but given the vehicle’s speed and the distance from
the cliff, it will certainly fall off the cliff and none of the passengers will
survive.
If the vehicle had been driven by a human driver who chose the first
rail and killed the child, no criminal liability would be imposed on him
owing to the general defense of necessity. An ai system can calculate the
probabilities for each option and choose the possibility with the lowest
number of casualties. Strong ai systems are already used to predict compli-
cated events (for example, weather patterns), and calculating the probabil-
ities in the locomotive example is considered much easier. The ai system
analyzing the possibilities is faced with the same two possibilities that the
human driver has to consider.
If the ai system takes into account the number of probable casualties,
based on its programming or as a result of the relevant machine learning,
it will probably choose to sacrifice the child. This choice by the ai system
meets all the conditions of necessity, so that if it were human, no criminal
liability would be imposed on it owing to the general defense of neces-
sity. Should the ai system be treated differently, then? Moreover, if the ai
system chooses the alternative option, and causes not the lesser but the
greater of the two evils to occur, we would probably want to impose crimi-
nal liability (on the programmer, the user, or the ai system) in exactly the
same way as if the driver were human.
Naturally, some moral dilemmas may be involved in these choices
and decisions: for example, whether it is legitimate for the ai system to
make decisions in matters of human life, and whether it is legitimate for
ai systems to cause human death or severe injury. But these dilemmas are
no different from the moral dilemmas of self-defense, discussed earlier.112
Moreover, moral questions are not to be taken into consideration in rela-
tion to the question of criminal liability. Consequently, it appears that the

general defense of necessity is applicable to ai systems in a similar way
to self-defense.
5.3.3. Duress
Can an ai system be considered as acting under duress in the criminal
law context? Duress is a justification of the same type as self-defense and
necessity. All are designed to mitigate the society’s absolute monopoly
on power, already discussed.113 The principal difference between self-
defense, necessity, and duress is in the course of conduct. In self-defense,
the defender’s reaction is intended to repel the attacker; in necessity, it is
a reaction against an external innocent object; and in the case of duress, it
is surrendering to a threat by committing an offense.
For example, a retired criminal with expertise in breaking into safes is
no longer active. He is invited by ex-friends to participate in another rob-
bery, where his expertise is required. He says no. They try to convince him,
but he still refuses. So they kidnap his daughter and threaten him that if
he does not participate in the robbery, they will kill her. He knows them
well, and knows that they are capable of doing it. He also knows that if he
involves police, they will kill his daughter. As a result, he surrenders to
the threat, participates in the robbery, and uses his expertise. If captured,
he can claim duress. Self-defense and necessity are irrelevant in this case
because he surrendered to the threat rather than face it.
The traditional approach toward duress is that under the right circum-
stances, it can justify the commission of offenses.114 The traditional justifi-
cation of the defense is the understanding in criminal law of the weakness
of human nature. At times, people would rather commit an offense under
threat than face the threat and pay the price by causing harm to precious
interests. Until the eighteenth century, the general defense of duress was
applicable to all offenses.115 Later, the Anglo-American legal systems nar-
rowed its applicability, and today it no longer includes severe homicide
offenses that require mens rea.116
In the safe- breaking example, if the offense were not robbery but
murder, the general defense of duress would not be applicable for the im-
position of criminal liability, but only as a consideration in punishment.
The reason for the narrow applicability is the sanctity of human life.117 But
there are many exceptions to this approach.118 In general, with the narrow
exception for homicide, duress enjoys broad application as a general de-
fense in criminal law worldwide. The individual who acts under duress is
considered to be choosing the lesser of two evils, from his or her point of
view: the evil of committing an offense by surrendering to the threat versus

the evil of harming a legitimate interest.
Given that the act performed under duress is the individual’s, carried
out when the authorities are not available or effective, the general de-
fense of duress partially reduces the applicability of the general concept
whereby only society (the state) has a monopoly on power, as discussed
earlier. Being in a situation that requires an act of duress is considered to
nullify the individual’s fault required for the imposition of criminal li-
ability. The modern basis of this defense is the understanding that an indi-
vidual may surrender to a threat rather than face it because not all people
are heroes, and no one is required by law to be a hero.119

are the life, freedom, body, and property of the individual or
of other individuals. No previous introduction between the
person threatened and the defender is required.120
3. The act of duress must be in surrender to a threat.
4. The act of duress must be proportional to the danger.122
5. The defender must not be in a situation to control the danger or
the conditions for its occurrence (actio libera in causa).123
If all these conditions are met, the individual is considered to be acting

under duress, and therefore no criminal liability is imposed on him or her
for the commission of the offense. But not every person surrendering to a
threat can be considered to act under duress; only when the act meets the
listed conditions in full. The question is whether the general defense of
duress is applicable to ai systems. The answer depends on the capabili-
ties of the ai systems to meet the conditions just outlined, in the relevant
situations.
Four of the listed conditions — numbers 1, 2, 4, and 5 — are almost
identical with those of self-defense and necessity, mutatis mutandis; the
difference is that in most legal systems, duress does not require an im-
mediate act because the threat and danger to the legitimate interest may
be continuous. The main difference between self-defense, necessity, and
duress lies in one of the conditions. In self-defense, the act is directed
toward the attacker; in the case of necessity, the act is directed toward an
external or innocent interest; and acting under duress is surrender to a
given threat.
The question concerning ai systems in this context is whether they are
capable of choosing the lesser of two evils. For example, a prison-guard

robot captures an escaping prisoner, who points a loaded gun at a human
prison guard and threatens that if he is not released immediately by the
robot, he will shoot the human guard. The robot calculates the probabili-
ties and determines that the danger is real. If the robot surrenders to the
threat, the guard’s life is saved, but an offense is committed and the robot
becomes an accessory to escape. If the robot does not surrender, no offense
is committed, the escape is foiled, but the human guard is killed.
If the guard who captured the prisoner were human, then no criminal
liability would be imposed on him owing to the general defense of duress
because all the conditions of this defense are met. An ai system can calcu-
late the probabilities for each option and choose the possibility with the
lowest number of casualties. Strong ai systems are already used to predict
complicated events (for example, weather patterns), and calculating the
probabilities in the prison-guard example is considered much easier. The
ai system analyzing the possibilities is faced with the same two possibili-
ties that the human guard has to consider.
If the ai system takes into account the number of probable casualties,
based on its programming or as a result of the relevant machine learn-
ing, it will probably choose to surrender to the threat. This choice by the
ai system meets all the conditions of duress, so that if it were human, no
criminal liability would be imposed on it owing to the general defense of
duress. Should the ai system be treated differently, then? Moreover, if the
ai system chooses the alternative option, and causes not the lesser but the
greater of two evils to occur, we would probably want to impose criminal
liability (on the programmer, the user, or the ai system) in exactly the same
way as if the prison-guard robot were human.
Naturally, some moral dilemmas may be involved in these choices and
decisions: for example, whether it is legitimate for the ai system to make
decisions in matters of human life, and whether it is legitimate for ai sys-
tems to cause human death or severe injury. But these dilemmas are no
different from the moral dilemmas of self-defense and necessity, discussed
earlier.124 Moreover, moral questions are not to be taken into consideration
in relation to the question of criminal liability. Consequently, it appears
that the general defense of duress is applicable to ai systems in a similar
way to self-defense and necessity.
5.3.4. Superior O rders

Is an ai system on official duty, committing an offense under orders from
its superior, protected from criminal liability? The general defense of su-
perior orders is relevant for individuals who discharge official duties in

authoritarian hierarchical official organizations such as the army, police,
and rescue forces. These individuals are often required to act against their
natural instincts. When faced with a gigantic fire, one’s natural instinct is
to escape, not to walk into it and save trapped people. One of the strongest
factors in the success of such missions is discipline. Disciplined soldiers
are likely to fulfill their mission even if this involves risking their lives.
Therefore, the first thing that soldiers are taught is discipline.
But at times, orders from superiors contradict criminal law, and carry-
ing them out involves committing an offense. There are two extreme models
for a solution to this problem. The first is the absolute defense model, in
which all actions under superior orders are protected.125 This model places
discipline above the rule of law, and only the commanders are to be held
criminally liable according to this model. The other extreme model is the
absolute responsibility model, in which no action is protected from crimi-
nal liability, even if performed under superior orders.126 This model places
the rule of law above discipline.
Under the first extreme model, the individual has no discretion in com-
mitting the offense. This model has led to the commission of war crimes
and crimes against humanity during World War II and other wars. Under
the second extreme model, the individual must be an expert in criminal
law or accompanied by an attorney at all times to make sure of not being
criminally liable for actions ordered by superiors. The disadvantages of
these models have led to the emergence of moderate models, the most
common of which is the manifestly illegal order model. According to this
model, the individual is protected from criminal liability unless he or she
performs a manifestly illegal order; if the individual performs an illegal
order that is not manifestly illegal, he or she is protected.127
The ultimate question concerns the difference between illegal and
manifestly illegal orders. Both types objectively contradict the law, but
the manifestly illegal order contradicts the public policy as well. Every so-
ciety follows a public policy that reflects its common values. These values
may be moral, social, cultural, religious, and so on. Different societies em-
brace different values and have differing public policies. A manifestly il-
legal order is one that harms these values, and thus harms society’s public
policy and its self-image.
For example, a soldier is ordered to rape a civilian who resists the mili-
tary operation. Rape is illegal, and in this situation it is manifestly illegal
because it violates the basic values of modern Western society. There is no
accurate and conclusive examination to distinguish between illegal and
manifestly illegal orders because public policy is dynamic, and it changes

with time, population, and social trends. But public policy may be taught,
usually inductively, case to case. Consequently, two main conditions must
be met for the general defense of superior orders to be applicable:
1. hierarchical subordination to authorized public authority; and

2. superior orders that require obedience and are not manifestly
illegal.
If both conditions are met, the individual is considered to be acting under

superior orders, and therefore no criminal liability is imposed on him or
her for the commission of the offense. But not every time an individual
obeys superior orders is the general defense applicable; only when the act
meets the listed conditions in full. The question is whether the general
defense of superior orders is applicable to ai systems. The answer to this
question depends on the capabilities of the ai systems to meet the condi-
tions just outlined, in the relevant situations.
The first condition has to do with the objective characteristics of the
relationships between the individual and the relevant organization.128
This relationship must involve a hierarchical subordination to authorized
public authority and a legitimate and operative system of hierarchical
order. Such systems exist in the army, police, and so on. Private organiza-
tions, however, have no authority to commit offenses. ai systems are in use
in many of these organizations, including the military, police, and prisons,
and are operated under superior orders for their regular activity. The sys-
tems perform various tasks under these orders.
The second condition has to do with the characteristics of the given
superior order. The order must require obedience; otherwise, it cannot
be considered an order. As far as its content is concerned, the order must
not be manifestly illegal. If the order is legal or illegal, but not manifestly
illegal, it satisfies this condition. Classification of the order as illegal or
manifestly illegal is at the discretion of the court, but it may be learned
inductively based on analysis of individual cases. ai systems equipped
with machine learning facilities are capable of inferring at least the general
outlines of manifestly illegal orders.129
For example, an ai drone operated by the air force is on a mission to
seek out a terrorist lab and destroy it. The drone finds its target, and then
delivers the information to headquarters. The information indicates that
the lab is populated by a known terrorist and his family. The drone re-
ceives the order to attack the lab using a heavy bomb. After calculating the
probabilities, the drone determines that all the people in the lab will die,
and then it executes the order. When the records of the drone are exam-
ined, it emerges that the drone understood the legality of the orders to be
in a grey area regarding the terrorist’s family, and that it could have flown
lower and destroyed the lab with fewer casualties.
If the drone were human, it would probably claim the general defense
of superior order, because international law accepts such orders under
certain situations, as long as the order is not manifestly illegal. Therefore,
a human pilot would probably have been acquitted in such a case, and
the general defense would have been applicable. Should an ai system be
treated differently? If both the human pilot and the ai system have the
same functional discretion, and both meet the conditions of this general
defense, there is no legitimate reason for using a double standard in such
cases.
Naturally, the criminal liability of the ai system, if any, does not affect
the criminal liability of its superiors, if any (in case of an illegal order).
Some moral dilemmas may also be involved in these choices and deci-
sions: for example, whether it is legitimate for the ai system to make deci-
sions in matters of human life, and whether it is legitimate for ai systems
to cause human death or severe injury. But these dilemmas are no different
from the moral dilemmas involving the applicability of other general de-
fenses, discussed earlier.130 Moreover, moral questions are not to be taken
into consideration in relation to the question of criminal liability. Conse-
quently, it appears that the general defense of superior orders is applicable
to ai systems.
5.3.5. De M inimis
Can the general defense of de minimis be applicable for ai systems? In
most legal systems, offenses are defined and formulated broadly, inevita-
bly resulting in over-inclusion or over-criminalization, so that cases that
are not supposed to be considered criminal end up being included within
the scope of the offense. At times, the criminal proceedings in these cases
would cause more social harm than benefit. For example, the case of a
fourteen-year-old girl who steals her brother’s basketball falls within the
scope of the offense of theft, and the question is whether it would be ap-
propriate to institute criminal proceedings for theft in such a case, consid-
ering its social consequences.
Most legal systems solve such problems by granting broad discretion to
the prosecution and the courts. The prosecution may decide not to initiate
criminal proceedings in cases of low public interest, and if such proceed-
ings are started, the court may decide to acquit the defendant owing to low
public interest. When the prosecution exercises this option, it is within its
administrative discretion. When the court exercises this discretion, it is
within its judicial power through the application of the general defense of
de minimis, which enables the court to acquit the defendant based on low
public interest in the case.
This type of judicial discretion has been widely accepted since ancient
times. Roman law, for example, determined that criminal law does not
extend to infant and petty matters (de minimis non curat lex), and that
the judge should not be troubled with such matters (de minimis non curat
praetor).131 In modern criminal law, the general defense of de minimis is
seldom exercised by the court given the wide administrative discretion
of the prosecution. But in some extreme cases, the court may exercise
this judicial discretion, in addition to the administrative discretion of the
prosecution.132
The basic test for de minimis defense is one of the social endangerment
involved in the commission of the offense. For the general defense of de
minimis to be applicable, the offense should reflect extremely low social
endangerment.133 Naturally, different societies at various times may attri-
bute different levels of social endangerment to the same offenses, because
social endangerment is dynamically conceptualized through morality, cul-
ture, religion, and so on. The relevant social endangerment is determined
by the court. The question is whether this general defense is relevant for
offenses committed by ai systems.
Applicability of de minimis defense depends on the particulars of the
case at hand and is based on its relevant aspects, and not necessarily on
the identity of the offender. The personality of the offender may also be
taken into consideration, but only as part of assessing the case. For this
reason, there is no difference between humans, corporations, or ai entities
regarding the applicability of this defense. The required low social endan-
germent is reflected in the factual event (in rem). For example, a human
driver swerves on the road and hits the guardrail. No damages are caused
to the guardrail or to other property, and there are no casualties. This is
a case for the de minimis defense, although it may be within the scope of
several traffic violations.
Would this case be legally different if the driver were not human but
an ai entity? And would it be different if the car were owned by a corpo-
ration? There is no substantive difference between humans, corporations,
and ai systems regarding the applicability of the de minimis defense, es-
pecially in light of the fact that this general defense is based on the char-
acteristics of the factual event and not necessarily on those of the offender.
Consequently, it appears that the general defense of de minimis is appli-

cable to ai systems.
C l osin g the Opening Ex ample: R obo t Killin g in Self- Defense

We opened this chapter with the example that described an ai robot, func-
tioning as a prison guard, who killed in self-defense. Returning to this
example, we can analyze it based on the insights gained from this chapter.
Let us assume that after examining the robot’s records, it transpires that
the robot had consolidated awareness of the standards of criminal law
regarding the factual components of the homicide. There is no doubt that
the human prison guard died as a result of being smashed against the wall
by the robot (factual causal connection). Following through this example
step by step, we may find several applicable general defenses.
First, the virus planted in the main computer interfered with the ro-
bot’s ability to create the right image of factual reality, and caused it to
ignore some of the rules of the system. Two types of general defenses may
be applicable as a result. The first is factual mistake, discussed previ-
ously. The reason for the factual mistake is immaterial for its legal con-
sequences, and therefore, if during the event the factual image of the
ai system differed from what actually happened because of the effect of
the virus, this can serve as the basis for the applicability of factual mistake
general defense.
The effect of the virus on the cognitive or volitive malfunction of the
robot during its operation should also be examined based on the records
of the ai system. If the virus caused the malfunction, then this can be the
basis for the applicability of another general defense having to do with loss
of self-control, insanity, or intoxication. The choice between these options
depends on the concrete effect of the virus on the ai robot. Because the
planting of the virus was involuntary (done by hackers), the consequences
of the effects of the virus will most likely prevent imposition of criminal
liability if the conditions of the relevant defense are met.
When the robot apprehended the two struggling humans, it misinter-
preted the situation as one in which two prisoners were struggling. This
misinterpretation can be proven by analyzing the robot’s records. If the
robot indeed misinterpreted the situation, whether or not as a result of
the virus, the general defense of factual mistake is applicable. The same
is true for the robot misinterpreting the command to keep away as if this
had been a threat to its mission. If the robot indeed interpreted this com-
mand as a threat, whether or not it was a threat in reality, can the reaction
of grabbing the human guard and smashing him against the wall be justi-
fied and defended?
As understood by the robot, this reaction is at most one of repelling a
threat. If all the conditions of self-defense are met, this may serve as the
basis for the general defense of self-defense. The robot’s mission, however,
was to prevent the escape of other prisoners but not to protect their lives,
freedom, body, or property. Therefore, self-defense is inapplicable in this
case, as are necessity and duress. Even if the system had orders to kill re-
sisting prisoners, the general defense of superior orders would not have
been applicable. Killing prisoners in response to an attempted escape, that
results in the escape of other prisoners (the disguised prisoner managed
to escape as a result of the robot’s actions), would likely be considered as
following a manifestly illegal order, obeying which is not protected.
But if it can be proven, based on the records, that the robot’s decision
to kill the guard was a direct consequence of the effect of the virus on the
system, then the legal situation may be different. If the virus caused the
robot to act automatically when killing the guard, or caused a malfunction
in its cognitive or volitive processes as a result of which the robot killed
the guard, then the general defenses of loss of self-control, insanity, or in-
toxication may be applicable according to the concrete effect of the virus.
Naturally, this does not reduce the criminal liability of others involved
in the event, including humans (superiors, prisoners, guards, and hackers),
corporations (the prison service and the manufacturer of the ai systems),
and other ai systems. It appears, however, that general defenses in criminal
law are applicable to ai systems as well as to humans and corporations.
6
Sen tencing ai
Is it possible to impose a prison sentence on an ai robot? How can such a

punishment be carried out in practice? One of the objectives of the crimi-
nal process is sentencing. The legal discussion taking place in court re-
garding the criminal liability of the defendant is often considered to be a
preliminary discussion before sentencing. Debating the criminal liability
of the defendant is not a theoretical exercise, but it has practical conse-
quences. If the defendant is acquitted, no punishment is imposed. But if
he or she is convicted, the punishment must reflect the attitude of society
toward the commission of the offense and toward the personal circum-
stances of the defendant.
The question is whether the sentencing process in criminal law and
criminal punishments are relevant to machines. For example, after all
the factual and mental element requirements of manslaughter were met
by an ai robot, the court convicted the robot of manslaughter, and the
prosecution and defense presented evidence for punishment. Under the
circumstances, if the robot had been a human, the court would have sen-
tenced this offender to three years of imprisonment. Is such a punishment
applicable to a robot? Is criminal punishment applicable to machines
at all?
This question is relevant and complementary to the examples that
began each of the previous chapters. To answer it, we must first explore
the general purposes of sentencing and their applicability to ai entities,
and then understand the legal technique of imposing punishments on non-
human entities. Finally, we will discuss the applicability of punishing
ai entities.
156
Sentencing AI . 157
6.1. C on ceptual Applicabili ty of C ri minal

Punish ment of ai Entities
6.1.1. ai P urpose of Sentencing:

C ombining R ehabilitation and I ncapacitation
Criminal punishments are imposed for a purpose. There are four common
general purposes of sentencing, which also function as the general sentencing
considerations: retribution, deterrence, rehabilitation, and incapacitation.
Retribution (lex talionis) is the most ancient purpose of sentencing.1 It
has to do with the past, and contains many aspects of revenge. The basic
idea of retribution is expressed in the maxim “an eye for an eye,”2 that
is, whatever you have caused shall be imposed on you. Retribution pur-
posefully causes suffering to the offender to a degree that fits the suffering
caused by the offender, and not more. Retribution functioned as a lenient
form of punishment because it prevented society from causing greater
suffering to offenders than they have caused to others.3 For example, the
death penalty would not be appropriate for breaking someone’s tooth.
The importance of retribution as a major objective of punishment grad-
ually diminished between the 1920s and the 1970s, primarily because
of criticism that retribution was archaic and retrospective, and therefore
did not sufficiently take into account the social benefit of punishment.4
The purpose of retribution was considered barbaric.5 Nevertheless, in light
of the actual and conceptual failure of rehabilitation, as we will discuss,
since the 1970s retribution has regained a significant position among the
purposes of punishment, in its modern form of “just desert.”6
Whereas rehabilitation concentrates on the personal characteristics
of the offender, thereby punishing the offender for his or her character,
retribution concentrates on the characteristics of the offense.7 The “just
desert” approach incorporates some personal aspects of the offender into
the retribution, thereby softening it.8 Thus, there are two important factors
in evaluating the offender’s punishment under retribution: (1) the factual
damage caused by the offender to society, and (2) the personal culpability
of the offender.9 Personal culpability is measured mostly by the mental
element of the offender.10
The question for our discussion is whether retribution is relevant to
ai systems. Retribution is intended primarily to satisfy society. Causing
suffering to the offender, in itself, has no prospective value. The suffering
may deter the offender, but that is part of the general purpose of deter-
rence, not of retribution. Retribution may have a cathartic effect on soci-

ety and the victims by causing the offender to suffer, but in this context,
punishing machines by retribution would be meaningless and impractical.
Some people, when they are in a great hurry and their car will not
start, may become angry and hit the car, kick it, or yell at it. Punishing
machines, including highly sophisticated ai robots, by retribution would
be the same as kicking a car. Some people may achieve a certain release of
their anger, but nothing more would be accomplished. Machines do not
suffer, and as long as retribution is based on suffering, it is not relevant to
punishing robots. This is true for both classic and modern (“just deserts”)
approaches to retribution.
Further, if retribution functions as a lenient form of sentencing aimed
at preventing revenge, its irrelevance to ai sentencing is even more pro-
nounced. Revenge is assumed to cause more suffering to the offender than
would the official punishment, but because machines do not experience
suffering, the distinction between revenge and retribution is meaningless
for them.
Deterrence is the teleological response to deontological retribution.
Deterrence emerged as the general purpose of sentencing toward the end of
the eighteenth century. Inspired by utilitarianism, legal scholars searched
for the prospective value of the law. This search included sentencing as
well, and it was assumed that criminal punishment may discourage of-
fenders and potential offenders from committing further offenses. Some
scholars called for the inclusion of punishment components as integral
parts of the definition of offenses.11 Deterrence reflected the human wish
for a better society, without delinquency.12
This better society was supposed to be achieved by intimidation. Of-
fenders and potential offenders, afraid of punishment, were expected to re-
consider their delinquent thoughts and eventually renounce their criminal
plan. Because punishment was supposed to deter people from committing
offenses, it was assumed that the more severe the punishment, the more
people would avoid committing offenses. But this assumption turned out
to be misleading for two reasons.
First, people are deterred more by efficient enforcement than by pun-
ishment.13 For example, the effect on driving habits of doubling the maxi-
mum penalty from $100 to $200 for exceeding the speed limit on a road
that has practically no police presence would be insignificant. Yet if more
police officers patrolled the same road, even if the fine were lowered to
$80, drivers would be more deterred from speeding. When city inspectors
are on strike and no citations are issued for parking in restricted areas,
Sentencing AI . 159
most people tend to park their cars in these areas, regardless of the level
of maximum punishment. Thus, the deterrent value of punishment alone
is very limited.14
Second, the deterrent value of punishment, even if it is maximal, tends
to erode in time with use, and at some point, the punishment may end
up functioning as an incentive for repeated offense.15 For instance, when
an offender is sentenced to prison for the first time, the deterrent effect
of the punishment is at its maximum level for that offender. The second
time, the effect is reduced, and the tenth time it is reduced dramatically.
After spending forty years in prison, the prisoner does not really have a
life outside of prison. Prison provides meals, social status, social security,
and friends. If released, the prisoner may commit offenses only to return
to this familiar social environment.
Thousands of homeless people in the Western world are committing
offenses on cold days to find shelter from the cold, have a warm meal,
and find a place to sleep. Not only does punishment not serve as a deter-
rent, but it serves as an incentive to repeat offending. In other words, the
punishment turns into the benefit of the offense. Nevertheless, for most
offenders and potential offenders, a combination of proper enforcement
with appropriate punishments can supply some deterrence, even if this is
not entirely effective. Our question, then, is whether deterrence is relevant
to ai systems.
Deterrence is intended to prevent the commission of the next offense,
by using intimidation. In the current technology, intimidation is a feeling
that machines cannot experience. Intimidation is based on a fear of future
suffering imposed if an offense is committed. Because machines do not
experience suffering, as we have already noted, both intimidation and the
reason for it are nullified when considering appropriate punishment for
robots. At the same time, both retribution and deterrence may be relevant
purposes for the punishment of human participants in the commission of
offenses by ai entities (for example, users and programmers).
Rehabilitation is a relatively newer purpose of modern sentencing,
based on the idea that the punishment and the process of sentencing may
serve as opportunities for correcting offenders’ social problems. Rehabili-
tation was inspired by the religious concept that every person may be
corrected given the right conditions, and has been embraced as a major
sentencing consideration since 1895 in Britain, spreading to the rest of
the world.16 The golden age of rehabilitation was between the 1920s and
1970s,17 and its popularity rested on its prospective quality that offered
hope for a better world.
Enthusiasm for rehabilitation waned with the publication of the Mar-

tinson report in 1974.18 The publication examined rehabilitation reports
in the state of New York and came to the surprising conclusion that “noth-
ing works.”19 The conclusions were reexamined repeatedly, and each time
confirmed,20 giving rise to various criticisms of rehabilitation, on the part
of research and even popular art.21 Disappointment with rehabilitation
triggered a revival of retribution, in its modern form as “just desert,” as
already noted. Rehabilitation was not abolished entirely, however, and
it is still being considered in the sentencing process, but more carefully.
The prospective purpose of rehabilitation is to address the roots of
delinquency. By solving the problems that lead to it, delinquency may be
solved, or at least dramatically reduced. To this end, the court must un-
derstand the causes of an offender’s delinquency, assess the potential for
rehabilitation (using such professionals as social workers and psycholo-
gists), and eventually decide on a proper punishment aimed at solving the
delinquency problem.22 One of the most popular trends in rehabilitation
is to grant the offender the cognitive and social skills needed to cope with
daily life.23
These tools are supposed to cause the offender to make inner changes
in relation to his or her social, family, or professional environment, provid-
ing an opportunity for resocialization and reintegration into society and
for leaving delinquency behind, to benefit the offender and also society.24
Our question is whether rehabilitation is relevant to ai systems.
ai systems can experience decision- making processes and reach de-
cisions that may appear unreasonable, as we have discussed.25 At times,
an ai system may need external guidance in order to refine its decision-
making process, which may be part of the machine-learning process. For
ai entities, rehabilitation functions in exactly the same way as for humans,
causing them to make better decisions in their everyday lives from the
point of view of society. Following this approach, the punishment of ai
entities would be directed at refining the machine-learning process.
Having been rehabilitated, an ai system would be able to form better
and more accurate decisions, by applying more limitations to its discretion
and by refining the process through machine learning. Thus, the punish-
ment, if correctly applied to an individual ai system, would become part of
the machine-learning process. Directed by the rehabilitative punishment,
the ai system would therefore have better tools to analyze factual data.
This is the same effect that rehabilitative punishment is intended to have
on humans by granting them better tools for facing factual reality.
Consequently, rehabilitation can be a relevant purpose of punishment
Sentencing AI . 161
for ai systems because it is not based on intimidation or suffering, but is

directed at improving the performance of the ai system. For humans, this
consideration may be secondary in many cases, but for ai systems it may
be a primary purpose of punishment. Nevertheless, rehabilitation is only
one of the considerations that are relevant to ai systems; incapacitation is
another.
Incapacitation is the physical prevention of offenders from committing
further offenses. This general consideration is based on skepticism toward
the effectiveness of the other prospective considerations (deterrence and
rehabilitation) to reduce delinquency rates; it focuses therefore on the
physical capabilities of delinquency. The basic approach is to neutralize
the offender’s physical capabilities to commit offenses.26 For example, in-
capacitation is the general consideration that justifies the chemical castra-
tion of sex offenders by eliminating their sexual impulse and their ability
to commit further sexual offenses.
For the physical prevention of delinquency, it is not necessary to
bring about inner changes or cognitive corrections in the offender’s mind.
The offender is not required to change his or her personality or values
for this consideration to be effective.27 In some ways, the consideration
of incapacitation is the expression of disappointment, even despair with
human nature, contrary to rehabilitation, which expresses hope. When
society feels hopeless toward an offender’s willingness or mental capabil-
ity to stop the delinquent activity, both deterrence and rehabilitation may
appear to be ineffective; but further offense must still be prevented, so
society resorts to incapacitation.
In this way, regardless of whether the offender understands or agrees
with the demands of society, the next offense is prevented because the
physical capability to commit the offense is neutralized. If in most cases,
rehabilitation is considered a lenient general consideration of punish-
ment, incapacitation is considered a harsh one. Our question is whether
incapacitation is relevant to ai systems.
If an ai system commits offenses and lacks the capability to change its
ways through machine learning, only incapacitation can supply an ad-
equate answer. Regardless of whether the ai system understands the mean-
ing of its activity, or whether the ai system is equipped with proper tools to
perform inner changes, delinquency must still be prevented. In a situation
of this type, society must disable the physical capabilities of the ai system
to commit further offenses, despite its other skills. This is what the society
does with human offenders in similar cases.28
We can conclude that two relevant considerations for the punishment
of ai systems are rehabilitation and incapacitation. Both reflect extreme

poles of sentencing, and both serve the purposes of criminal law regard-
ing nonhuman offenders. When the ai system is capable of carrying out
inner changes that affect its activity, rehabilitation appears to be preferable
to incapacitation, but when it does not possess such capabilities, inca-
pacitation is preferable. Thus, the punishment is adjusted to the relevant
personal characteristics of the offender, as in the case of human offenders.
6.1.2. The L egal Technique of C onversion:

The C ase of C orporations (R ound 2)
We have seen that sentencing considerations are relevant to ai systems.
The question is, how can we impose human punishments on machines?
For example, how can we impose imprisonment, fines, or capital punish-
ment on ai systems? To do so, we need a legal technique of converting
human penalties into penalties for ai entities. Such technique may be in-
spired by the one used to convert human penalties into penalties suitable
for corporations. The first time we examined corporations (Round 1), we
were concerned with the idea of imposing criminal liability on ai sys-
tems.29 This time (Round 2) we are concerned with the idea of imposing
punishments on these systems.
Not only has criminal liability been imposed on corporations for cen-
turies, but corporations have also been sentenced, and not only to fines.
Corporations are punished in various ways, including imprisonment. Note
that corporations are punished separately from their human officers (di-
rectors, managers, employees, and so on) in exactly the way that criminal
liability is imposed on them separately from the criminal liability, if any,
of their human officers. There is no debate over the question whether cor-
porations should be punished using a variety of punishments, including
imprisonment — the question concerns only how to do it.30
To answer this question of “how,” a general legal technique of conver-
sion is needed, as just noted. This operation is carried out in three prin-
cipal stages. First, the general punishment itself (for example, imprison-
ment, fine, probation, or death) is analyzed regarding its roots of meaning.
Second, these roots are sought in the corporation. Third, the punishment
is adjusted according to the roots found in the corporation. For example,
in the case of imposition of incarceration on corporations, first incarcera-
tion is traced back to its roots in the act of depriving individuals of their
freedom, and then a meaning is sought for the concept of freedom for
corporations.
After this meaning has been understood, in the third and final stage,
Sentencing AI . 163
the court imposes a punishment that is the equivalent of depriving a cor-

poration of its freedom. This is how the general legal technique of conver-
sion works in the sentencing of corporations. At times, this requires the
court to be creative in the adjustments needed to make punishments ap-
plicable to corporations, but the general framework is clear and workable,
and it has been implemented with all types of punishments imposed on
all types of corporations.31
An excellent example is the case of the Allegheny Bottling Company,32
a corporation found guilty of price-fixing (antitrust). It was agreed that
under the given circumstances, if the defendant were human, the appro-
priate punishment would be imprisonment for a certain term. The ques-
tion concerned the applicability of imprisonment to corporations — in
other words, a question of “how.” As a general principle, the court de-
clared that it “does not expect a corporation to have consciousness, but it
does expect it to be ethical and abide by the law.”33
The court did not find any substantive difference between humans
and corporations in this matter, adding that “[t]his court will deal with
this company no less severely than it will deal with any individual who
similarly disregards the law.”34 This statement reflects the basic principle
of equalizing punishments of human and corporate defendants.35 In this
case, the corporation was sentenced to three years’ imprisonment, a fine of
$1 million, and probation for a period of three years. The court proceeded
to discuss the idea of corporate imprisonment based on the three stages
just described.
First, the court asked what were the general meanings of imprison-
ment, and accepted the definitions of imprisonment as “constraint of a
person either by force or by such other coercion as restrains him within
limits against his will” and as “forcible restraint of a person against his
will.” The court’s conclusion was simple and clear: “[t]he key to corpo-
rate imprisonment is this: imprisonment simply means restraint” and
“restraint, that is, a deprivation of liberty.” The court’s conclusion was
reinforced by several provisions of the law and also of case laws. Conse-
quently, “[t]here is imprisonment when a person is under house arrest, for
example, where a person has an electronic device which sends an alarm if
the person leaves his own house.”
This ended the first stage. In the second stage, the court searched for
a meaning of this punishment for corporations, concluded that “[c]orpo-
rate imprisonment requires only that the Court restrain or immobilize the
corporation,”36 and proceeded to implement the prison sentence on the
corporation according to this insight. Thus, in the third and final stage,
the court made imprisonment applicable to the corporations and imple-

mented it as follows:
Such restraint of individuals is accomplished by, for example,

placing them in the custody of the United States Marshal. Like-
wise, corporate imprisonment can be accomplished by simply
placing the corporation in the custody of the United States Mar-
shal. The United States Marshal would restrain the corporation
by seizing the corporation’s physical assets or part of the assets
or restricting its actions or liberty in a particular manner. When
this sentence was contemplated, the United States Marshal for the
Eastern District of Virginia, Roger Ray, was contacted. When asked
if he could imprison Allegheny Pepsi, he stated that he could.
He stated that he restrained corporations regularly for bankruptcy
court. He stated that he could close the physical plant itself and
guard it. He further stated that he could allow employees to come
and go and limit certain actions or sales if that is what the Court
imposes.
Richard Lovelace said some three hundred years ago, “stone
walls do not a prison make, nor iron bars a cage.” It is certainly true
that we erect our own walls or barriers that restrain ourselves. Any
person may be imprisoned if capable of being restrained in some
fashion or in some way, regardless of who imposes it. Who am I
to say that imprisonment is impossible when the keeper indicates
that it can physically be done? Obviously, one can restrain a corpo-
ration. If so, why should it be more privileged than an individual
citizen? There is no reason, and accordingly, a corporation should
not be more privileged.
Cases in the past have assumed that corporations cannot be
imprisoned, without any cited authority for that proposition. . . .
This Court, however, has been unable to find any case which ac-
tually held that corporate imprisonment is illegal, unconstitu-
tional or impossible. Considerable confusion regarding the ability
of courts to order a corporation imprisoned has been caused by
courts mistakenly thinking that imprisonment necessarily involves
incarceration in jail. . . . But since imprisonment of a corporation
does not necessarily involve incarceration, there is no reason to
continue the assumption, which has lingered in the legal system
unexamined and without support, that a corporation cannot be
imprisoned. Since the Marshal can restrain the corporation’s lib-
Sentencing AI . 165
erty and has done so in bankruptcy cases, there is no reason that

he cannot do so in this case as he himself has so stated prior to the
imposition of this sentence.37
Thus, imprisonment can be applied not only to human but also to

corporate offenders. Following the same approach, imprisonment is not
the only penalty applicable to corporations, but other penalties can be
converted as well, even if they were originally designed for human offend-
ers. And if this is true for imprisonment, which is an essentially human
penalty, fine can be easily collected from corporations, just like taxes. This
raises the equivalent question regarding ai systems.
Using the general legal technique of conversion, as just presented and
applied to corporate delinquency, human punishments can be made ap-
plicable in the same way and imposed on ai systems as well. Next we ex-
amine the applicability of common punishments in modern criminal law
to ai systems.
6.2. Applicabili ty of Punish ments to ai Entities

We will review the applicability and imposition of the following punish-
ments on ai systems: capital punishment, imprisonment, probation, public
service, and fine.
6.2.1. C apital P unishment

The death penalty is one of the most ancient punishments in human his-
tory, and it is considered the most severe penalty in most cultures.38 In the
past it was a common penalty, but since the eighteenth century, the global
approach has been to restrict and minimize its use. Therefore, capital pun-
ishment has been replaced by more lenient penalties for many offenses,
and methods of execution were developed to cause minimal suffering to
the offenders.39 For example, in the eighteenth century, the length of the
noose was increased to cause a faster and less painful death, and the guil-
lotine was introduced for the same reason.
Cruel methods of execution were prohibited. For example, dismember-
ing by tying the offender’s arms and legs to running horses was prohibited
because of the suffering it caused. Some countries abolished capital pun-
ishment altogether, but most countries did not. The United States Supreme
Court ruled in 1979 that capital punishment, in the appropriate cases, does
not infringe on the Eighth Amendment to the Constitution (that is, it is not
cruel and unusual punishment), and therefore it is constitutionally valid.40
The methods of execution used in the United States were not considered
unconstitutional either.41
Retribution justifies capital penalty only for severe offenses in which
penalty parallels the offense in suffering or result (for example, homicide
offenses).42 Deterrence can justify capital punishment only as a deterrent
of the public and not of the offender, because a dead person cannot be
deterred.43 Rehabilitation is completely irrelevant for this punishment, as
dead people cannot be rehabilitated. And given that retribution and de-
terrence are irrelevant for ai sentencing, as noted previously, and because
rehabilitation is irrelevant to capital punishment, the only general consid-
eration that supports capital punishment in the sentencing of ai entities
is incapacitation.
There is no doubt that a dead person is incapacitated from commit-
ting further offenses and that the most dominant consideration for the
death penalty is incapacitation.44 The question regarding the applicability
of capital punishment to ai systems is how to impose the death penalty
on them. To answer this question, we must follow the three-stage model
presented earlier.
First, we should analyze the meaning of capital punishment; second,
seek out its roots as far as ai systems are concerned; and finally, adjust the
punishment to these roots. Functionally, capital punishment is depriva-
tion of life. Although this deprivation may affect not only the offender but
other people as well (relatives, employees, and so on), the essence of the
death penalty is the death of the offender. When the offender is human, life
means the person’s existence as a functioning creature. When the offender
is a corporation or an ai system, its life may be defined through its activity.
A living ai system is a functioning one, therefore the “life” of an ai
system is its capability to function as such. Stopping the ai system’s ac-
tivity does not necessarily mean the “death” of the system. Death means
permanent incapacitation of the system’s life. Therefore, capital punish-
ment for an ai system means its permanent shutdown so that no further
offenses or any other activity can be expected on the part of the system.
When an ai system is shut down by court order, this means that the society
prohibits the operation of that entity because it is considered too danger-
ous for society.
This application of capital punishment on ai systems serves both the
purposes of capital punishment and of incapacitation (as a general pur-
pose of sentencing) in relation to ai systems. When the offender is too
dangerous for society, and society decides to impose the death penalty, if
the punishment is acceptable in the given legal system, it is intended to
Sentencing AI . 167
accomplish the total and final incapacitation of the offender. This is true
of human offenders, corporations, and ai systems. For ai systems, perma-
nent incapacitation means absolute shutdown under court order, with no
option of reactivating the system again.
A permanently incapacitated system will not be involved in further
delinquent events. It may be argued that such shutdown may affect other
innocent persons (for example, the manufacturer of the system, or its pro-
grammers and users). Yet this is true not only for ai systems, but for human
offenders and corporations as well. The execution of an offender also af-
fects his innocent family (in the case of a human offender) or affects em-
ployees, directors, managers, shareholders, and so on (in the case of a cor-
poration). When the offender is an ai system, shutdown also affects other
innocent persons, but this is not unique to ai systems. Therefore, capital
punishment may be applicable to ai systems.
6.2.2. I mprisonment and Suspended I mprisonment

Imprisonment is a general term for various penalties. The common charac-
teristic of these penalties is deprivation of the offender’s liberty. Physical
incarceration is only one type of imprisonment, although it is the most
common. For example, under certain circumstances, public service can
function as imprisonment. The purposes of imprisonment have differed
in different societies and times. At times, the purpose was to make the of-
fender suffer; at other times, it was to rehabilitate the offender under tight
discipline.45 But since the eighteenth century, a dominant consideration
for the evaluation of imprisonment has been its social efficiency.46
The social efficiency of the imprisonment is evaluated by the rate of
recidivism. If the rate of recidivism remains constant or increases, then im-
prisonment is not considered socially efficient. Therefore, the imposition
of imprisonment has been used to initiate an inner change in the offender
for the benefit of society.47 Thus, prisoners have been taught professions;
have learned reading and writing; have been helped to break drugs, alco-
hol, and violent habits; and have been encouraged to experience working
during their stay in prison. Imprisonment has been adapted to fit various
populations, resulting in the creation of such solutions as supermax pris-
ons48 and shock incarceration.49
Nevertheless, because imprisonment has become a popular penalty,
prison overcrowding is currently considered its most acute problem. Retri-
bution supports imprisonment, especially incarceration, because it makes
the offender suffer. Deterrence supports imprisonment because the suffer-
ing caused in prison may deter offenders from recidivism and potential
offenders from offending. Both retribution and deterrence, however, are ir-
relevant to ai systems because they experience neither suffering nor fear.50
Imprisonment for ai systems should be evaluated based on rehabilitation
and incapacitation, which are relevant to ai entities.
When the offender is deprived of his or her liberty, society can use this
prison term for the offender’s rehabilitation, by initiating inner change.
This change can be the result of activity carried out in prison. If the of-
fender accepts the inner change and does not return to delinquency, then
imprisonment is considered successful and the offender rehabilitated. At
the same time, when the offender is under strict supervision inside the
prison, his or her capability for committing further offenses is dramatically
reduced, which may be considered incapacitation as long as supervision
prevents recidivism.
Here is the question regarding the applicability of imprisonment: how
can imprisonment be imposed on ai systems? To answer this question,
we again follow the three-stage model. First, we analyze the roots of the
meaning of imprisonment; second, we seek the meaning of these roots for
ai systems; and finally, we adjust the punishment to the roots we have re-
vealed for ai systems.
Functionally, imprisonment is deprivation of liberty. Although this
deprivation may affect not only the imprisoned offender, but other people
as well (for example, relatives, employees, and so on), the essence of the
imprisonment consists of restricting the offender’s activity. When the of-
fender is human, liberty means the person’s freedom to act in any desired
way. When the offender is a corporation or an ai system, its liberty may
also be defined through its activity. The freedom of an ai system lies in the
exercise of its capabilities without restriction, including both the exercise
of these capabilities and their content.
Thus, the imposition of imprisonment on ai systems consists of de-
priving them of their liberty to act by restricting their activities for a spec-
ified period of time and under strict supervision. During this time the
ai system may be repaired to prevent the commission of further offenses.
Repair of the ai system may be more efficient if the system is incapacitated,
especially if it is done under court order. This situation can serve both
purposes of rehabilitation and incapacitation, which are the relevant sen-
tencing purposes for ai systems. When the ai system is in custody, under
restriction and supervision, its capability to offend is incapacitated.
An ai system being repaired through inner changes, initiated by ex-
ternal factors (for example, programmers working under court order), and
gaining experience during the period of restriction, can be substantively
Sentencing AI . 169
considered to be in rehabilitation because the chances of the system being

involved in further delinquency are being reduced. The social value of
imposing imprisonment on ai systems is real, as the dangerous system is
being removed from society for repairs, and at the same time, it is inca-
pable of causing further harm to society. When the process is complete, the
system may be returned to full activity. If the system is beyond rehabilita-
tion, then incapacitation becomes the principal objective, dictating a long
period of imprisonment or even capital punishment.
A sentence of suspended imprisonment is a conditional penalty.51 The
offender is warned that if further offense is committed, the full penalty
of imprisonment is imposed for the new offense, and in addition, the of-
fender will have to serve another term of imprisonment for the original of-
fense. This penalty is intended to deter the offender from offending again,
for at least as long as the condition remains in force. This penalty is im-
posed by adding the relevant line to the offender’s criminal record. The
question of suspended sentences does not concern the execution of the
penalty, but concerns its social meaning when imposed on ai systems.
A sentence of suspended imprisonment for an ai system is an alert to
reconsider its course of conduct. This process may be initiated by pro-
grammers, users, or the manufacturer, in the same way that human of-
fenders may be assisted by their relatives or professionals (for example,
psychologists or social workers), and corporate offenders by their officers
or professionals. A suspended sentence is a relatively lenient measure
calling for reconsideration of one’s course of conduct. In essence, it is not
substantially different from imprisonment, although for human offenders
it may be vastly different because it spares the human offender the suf-
fering of incarceration and deters him or her from delinquency through
intimidation.
6.2.3. P robation
Prison overcrowding has forced states to develop substitutes for incar-
ceration. One of the popular substitutes is probation, developed in the
mid-nineteenth century by private charity and religious organizations that
undertook to care for convicted offenders by giving them the social tools
they needed to abandon delinquency and reintegrate into society.52 Most
countries embraced probation as part of their public sentencing systems.
The first US state to do that was Massachusetts, in 1878. Thus, probation
was operated and supervised by the state.53
The social tools provided to the offender vary from case to case. When
the offender is addicted to drugs, the social tools include a drug rehabili-
tation program. When the offender is unemployed or has no profession,

the social tools include vocational training. During the time the offender
is under probation, the authorities supervise and monitor him or her to
ensure that no further offenses are committed. Probation is dominantly
rehabilitative, and is appropriate for offenders who have a high potential
for rehabilitation. Consequently, the court needs an accurate assessment
of that potential, prepared by the probation service, before sentencing the
offender to probation.54
Retribution is irrelevant for probation, because probation is not in-
tended to make the offender suffer. Neither is deterrence relevant to pro-
bation, which is perceived as a lenient penalty with a negligible deterrent
value. Moreover, as we have already seen, both retribution and deterrence
are irrelevant to ai sentencing. Incapacitation, which is relevant to ai sen-
tencing, is not reflected in probation, unless the framework for the specific
probation is extremely strict and the offender’s delinquent capabilities are
incapacitated. The dominant purpose of probation, however, is rehabilita-
tion through the provision of the relevant social skills needed to reinte-
grate the offender into society.
Here is the question regarding the applicability of probation: how can
probation be imposed on ai systems? To answer this question, we again
follow the three-stage model. First, we analyze the roots of the meaning
of probation; second, we seek the meaning of these roots for ai systems;
and finally, we adjust the punishment to the roots we have revealed for
ai systems.
Functionally, probation consists of supervising the offender and pro-
viding him or her with the means to reintegrate into society. These mea-
sures should match the type of delinquency that was the immediate cause
for the sentencing.55 Probation works as a functional correction of the of-
fender. When an offense is committed by an ai system, the system must be
diagnosed to determine whether it can be corrected. At this stage, human
offenders are diagnosed to determine their potential for rehabilitation.
Both types of diagnosis are performed by professionals.
Human offenders may be diagnosed by probation service staff, social
workers, psychologists, psychiatrists, physicians, and so on. ai system of-
fenders may be diagnosed by technology experts. If the diagnosis shows
no potential for rehabilitation, then the ultimate purpose of sentencing
becomes incapacitation, because society wishes to prevent further harm.
But if the offender has a high potential for rehabilitation, then proba-
tion and rehabilitation as the purpose of sentencing are taken into con-
sideration. This is equally true for human offenders, corporations, and
Sentencing AI . 171
ai systems. The core question for ai system diagnosis is whether the

system may be repaired in its current configuration, for example, through
machine learning.
If the conclusion is that the appropriate penalty is probation, it must
be applied with reference to the problems that the delinquency raised.
This is true for all types of offenders. The objective of the treatment is to
fix a certain problem, therefore the treatment should match the problem.
The difference between probation and imprisonment, which also includes
treatment, is the incapacitation of the offender’s delinquent capabilities
during the period of the penalty. If the offender is not considered danger-
ous during the period of the treatment, probation may be suitable, but if
the offender must be incapacitated during treatment, imprisonment may
be better suited.
Therefore, if society finds that the ai system can continue function-
ing while under treatment, probation may be suitable. If it is too dan-
gerous for society, imprisonment may be suitable. When probation is fi-
nally imposed, the ai system must begin the treatment, which consists of
repair of its inner processes. Some systems may require intervention in the
machine-learning process, others require upgrades of their hardware, and
still others require intervention in the basic software of the system. During
this process, the ai system continues its routine activity under supervision
imposed by the court.
Socially and functionally, probation is identical for human offenders,
corporations, and ai systems. The attributes of each offender require dif-
ferent treatment, but that is true with respect to human offenders as well.
Naturally, the manufacturers, programmers, and users of the system can
initiate the repair process without the intervention of the court, but when
the court orders the intervention, it means that this is the will of society.
In the same way, a drug addict or the addict’s family may initiate a drug
rehabilitation process without a court order, but when it is imposed by the
court, it signifies that this is the will of society.
6.2.4. P ublic Service

Public or community service is a substitute for imprisonment developed
to ease prison overcrowding. For offenses that are not severe, the court
may impose public service on the offender instead of other penalties, or
in addition to them.56 The offender is not incapacitated by this penalty,
but on the contrary is forced to contribute to society in “compensation”
for his or her involvement in the delinquency. In this way, society signi-
fies that the delinquency is unacceptable, but because the social harm is
not severe, lenient measures are taken to bring about the required inner
change in the offender.
Public service has yet another dimension, which relates to the com-
munity. Public service is carried out in the offender’s community to signify
that he or she is part of that community, and that causing harm to the com-
munity boomerangs back onto the offender.57 In many cases, public service
is added to probation in order to improve the chances of the offender’s full
rehabilitation within the community.58 Public service has more than mere
compensational value, as it proposes to make the offender understand and
become sensitive to the needs of the community. Public service is part of
the learning and reintegration processes that the offender experiences.
Retribution is irrelevant for public service, because public service is
not intended to make the offender suffer. Neither is deterrence relevant
for public service because it is perceived as a lenient penalty with neg-
ligible deterrent value. Both retribution and deterrence are irrelevant to
ai sentencing. Incapacitation, which is relevant for ai sentencing, is not
reflected in public service, unless the framework for the specific public
service is extremely strict and the offender’s delinquent capabilities are
incapacitated.
The dominant purpose of public service, however, is rehabilitation
through reintegration of the offender into society. Here is the question re-
garding the applicability of public service: how can public service be im-
posed on ai systems? To answer this question, we again follow the three-
stage model. First, we analyze the roots of the meaning of public service;
second, we seek the meaning of these roots for ai systems; and finally, we
adjust the punishment to the roots we have revealed for ai systems.
Functionally, public service consists of supervised compensation to
society provided by the experience of integration with society. The of-
fender expands his or her experience with society, making integration
easier. Broadening the offender’s social experience benefits society be-
cause it also includes a compensational dimension. Social experience is
not exclusive to human offenders. Both corporations and ai systems have
strong interactions with the community. Public service may empower and
strengthen these interactions, and make them the basis for the required
inner change.
For example, a medical expert system, equipped with machine-
learning capabilities, is used in a private clinic to diagnose patients. The
system was considered negligent, and the court has imposed public ser-
vice on it. To carry out this penalty, the system may be used by public
medical services or public hospitals. This can serve two main goals. First,
Sentencing AI . 173
the system is exposed to more cases, and through machine learning it can
refine its functioning. Second, its work can be considered compensation
to society for the harm caused by the offense.
At the end of the public service term, the ai system is more experi-
enced, and if the machine-learning process was effective, system perfor-
mance is improved. Because public service is supervised, regardless of
whether it is accompanied by probation, the machine-learning process
and other inner processes are directed toward the prevention of further
offenses during the public service. By the end of the public service period,
the ai system will have contributed time and resources to the benefit of so-
ciety, and this may be considered compensation for the social harm caused
by the commission of the offense. Thus, the public service of ai systems
resembles human public service in its substance.
It may be argued that the compensation is in practice contributed by
the manufacturers or users of the system, because they are the ones who
suffer from the absence of the system’s activity. This is true not only for
ai systems, but also for human offenders and corporations. When a human
offender performs public service, his or her absence is felt by family and
friends. When a corporation carries out public service, its resources are
not available to its employees, directors, clients, and so on. This absence
is part of carrying out the public service, regardless of the identity of the
offender, so that ai systems are not unique in this regard.
6.2.5. F ine
Fine is a payment made by the offender to the state treasury. It is not con-
sidered compensation because it is not given directly to the victims of the
offense, but to society as a whole. If we regard society as the victim of any
offense, we may view a fine as a type of general compensation. The fine
has evolved from the general remedy of compensation, when the criminal
process was still between two individuals (private plaintiff versus defen-
dant).59 When criminal law became public, the original compensation was
converted into a fine. Today, criminal courts may impose both fines and
compensations as part of the criminal process.
In the eighteenth century, the fine was not considered a preferred
penalty because it was not deemed to be as strong a deterrent as impris-
onment.60 But during the twentieth century, when prisons became over-
crowded and the cost of maintaining prisoners in state prisons increased,
the penal system was urged to increase the use of fines.61 To facilitate the
efficient collection of fines, in most legal systems the court can impose
imprisonment, public service, or confiscation of property in case of non-
payment.62 The fine imposed is not necessarily proportional to the harm

caused by the offense, but to the severity of the offense.
Retribution may be relevant to fines, if the fine is proportional to the
social harm and reflects it. Deterrence may also be relevant to fines, if the
fine causes a loss of property sufficiently great to deter. But as noted ear-
lier, both retribution and deterrence are irrelevant to ai sentencing. Fines
have no prominent rehabilitative value, although paying the fine may re-
quire additional work from the offender, allowing less free time to commit
offenses. Here is the question regarding the applicability of fines: how can
fines be imposed on ai systems? The main difficulty is that ai systems pos-
sess no money or other property of their own.
Corporations possess property, so paying fines is the easiest way of
imposing a penalty on corporations, but ai systems do not possess prop-
erty. To answer our question, we again follow the three-stage model. First,
we analyze the roots of the meaning of fines; second, we seek the meaning
of these roots for ai systems; and finally, we adjust the punishment to the
roots we have revealed for ai systems. The result should also include ap-
propriate solutions for ineffective fines, that is, cases in which there are
difficulties in collecting the fines.
Functionally, a fine is a forced contribution of valuable property to
society. In most cases a fine is expressed as money, but in certain legal
systems, other valuable property may be used. In some legal systems, the
amount of the fine is determined based on the cost of a work day, week, or
month of the defendant, with the fine matching this cost.63 Even if the fine
is determined as an absolute sum, in most cases the absence of this sum
in the offender’s pocket translates into the work hours needed to produce
the missing amount. Therefore, a fine generally reflects work hours, days,
weeks, or months, depending on the amount and on the offender’s wealth
or skills.
As we discussed earlier in the context of public service, the produc-
tivity of an ai system can also be evaluated in work hours for the com-
munity.64 It is true that ai systems do not possess property, but they are
capable of working, which is valuable and may be assigned a monetary
value. For example, the work hour of a medical expert system may be as-
sessed as $500. If a fine of $1,000 is imposed on an ai system, the fine can
be translated into two work hours of the system. Therefore, the system can
pay by using the only currency it possesses: work hours.
The work hours are contributed to society, in the same way that public
service is contributed. When a human offender does not have the required
sum of money to pay the fine, other penalties are imposed. One of these
Sentencing AI . 175
is public service, which may be measured by the number of work hours.

Using work hours as a measure of payment can serve not only the pur-
pose of paying the fine (by an accurate accounting of the number of hours
worked), but also the optional purpose of enforcing the fine together with
public service, imprisonment, or any other relevant penalty.
It may be argued that this payment to society is in practice contributed
by the manufacturers or users of the ai system, because they are the ones
who suffer from the absence of the system’s activity while it is working to
pay the fine. This is true not only for ai systems, but also for human offend-
ers and corporations. When a human offender pays a fine, the absence of
the money (or of the person, if additional work hours are needed to make
up for the absence of money) is felt by family and friends. When a corpora-
tion pays a fine, its resources are not available to its employees, directors,
clients, and so on. This absence is part of paying the fine, regardless of the
identity of the offender, so that ai systems are not unique in this regard.
C l osin g the Opening Ex ample: Sentencing the Killin g R obo t

We opened this chapter with an example that described an ai robot con-
victed of manslaughter. We can now return to this example and analyze
it based on the insights gained from this chapter. After the court has con-
victed the ai robot for manslaughter, the criminal process continues as if
the offender were human, and the next stage is that of sentencing. The
prosecution and the defense may argue for a harsh or lenient penalty, and
eventually the court decides the appropriate penalty based on the relevant
circumstances of the case.
In our example, after assessing the circumstances, the court reaches the
conclusion that if the robot were human, the relevant penalty would be
three-year prison sentence. If the robot had been a corporation, the court
would have imposed three years of imprisonment, which would have been
carried out as demonstrated earlier.65 The same should be the case with
the ai robot. Therefore, if the judicial discretion of the court results in the
imposition of three years’ imprisonment, this is the penalty that must be
imposed, regardless of whether the defendant is a human, a corporate, or
an ai offender.
After its imposition, the penalty must be carried out. As previously
noted, the imposition of imprisonment on ai systems takes the form of
depriving them of their liberty to act by restricting their activity for the
required term and under strict supervision. During this time, the ai system
can be repaired to prevent the commission of further offenses. Repair of
the ai system may be more efficient if the system is incapacitated and

when it happens under court order. This situation serves both the pur-
poses of rehabilitation and incapacitation, which are the relevant sentenc-
ing purposes in the case of ai systems. When the ai system is in custody,
restricted, and supervised, its capabilities to offend are incapacitated.
Repairing the ai system through inner changes, initiated by external
factors (for example, programmers working under court order) and experi-
enced by the system as a restriction, is substantive rehabilitation because
the chances of the system being involved in further delinquency are re-
duced. The social value of imposing imprisonment on an ai system is real.
The dangerous system is being removed from society for repairs, during
which time it is not capable of causing further harm to society. When the
process is complete, at the end of the term determined by the court, the
system can be returned to full activity, the same way that a prisoner who
finishes serving a penalty is returned to society.
C on clusion s
Criminal liability for artificial intelligence entities may sound radical. For
centuries, criminal liability was considered to be part of an exclusively
human universe. The first crack in the concept occurred in the seventeenth
century, when corporations were admitted into this exclusive club. Cor-
porate offenders joined human offenders, as both criminal liability and
punishments were equally imposed on both. The twentieth century intro-
duced ai technology to humankind. This technology developed rapidly in
an attempt to imitate human capabilities. Today, ai systems imitate some
human capabilities perfectly; indeed, they outperform humans in many
areas. Other human capabilities, however, still cannot be imitated.
Criminal liability does not require that offenders possess all human
capabilities, only some. If an ai entity possesses these capabilities, then
logically and rationally, criminal liability can be imposed whenever an
offense is committed. The idea of criminal liability of ai entities should
not be confused with the idea of moral accountability, which is a hugely
complex issue not only for machines, but for humans. Morality, in general,
has no single definition that is acceptable to all societies and individuals.
Deontological and teleological morality are the most acceptable types,
and in many situations they lead in opposite directions, both “moral.” The
Nazis considered themselves to be deontologically moral, although most
other people, societies, and individuals thought otherwise. Because moral-
ity is so difficult to assess, moral accountability is not the most appropriate
and efficient way of evaluating responsibility in criminal cases. Therefore,
society has chosen criminal liability as the most appropriate means for
dealing with delinquency. Modern criminal liability is independent from
morality of any kind, as well as from the concept of evil. It is imposed in
an organized, almost mathematical way.
Criminal liability has definite requirements, not more and not less. If
an ai entity meets all these requirements, then there is no reason not to
impose criminal liability on it. We have seen that ai entities are capable
of meeting these requirements in the same way that human and corporate
offenders do (but not animals). We have also seen that all types of crimi-
177
nal liability are relevant to ai technology, and all substantive arguments of

criminal liability can be made regarding ai entities. Finally, we have also
seen that penalties can be imposed on ai entities in a way that substan-
tively resembles the way in which corporations are being punished.
Society derives the usual benefits of criminal law by imposing crimi-
nal liability on ai entities. The conclusion is simple. Either we impose
criminal liability on ai entities, or we must change the basic definition of
criminal liability as it developed over thousands of years, and abandon the
traditional understandings of criminal liability.
Note s
Preface
1. This is the opening example of chapter 2, analyzed there in detail.
1. The Emergence of Machina Sapiens Criminalis

1. See, for example, aage gerhardt drachmann , the mechanical tech -
nology of greek and roman antiquity : a study of the literary sources
(1963); j . g . landels , engineering in the ancient world (rev. ed. 2000).
2. rené descartes , discours de la méthode pour bien conduire sa raison
et chercher la vérité dans les sciences (1637) (Discourse on the Method of
Rightly Conducting One’s Reason and of Seeking Truth in the Sciences).
3. Terry Winograd, Thinking Machines: Can There Be? Are We?, in the
foundations of artificial intelligence 167, 168 (Derek Pertridge & Yorick
Wilks eds., 1990, 2006).
4. thomas hobbes , leviathan or the matter , forme and power of a
common wealth ecclesiasticall and civil III.xxxii.2 (1651): “When a man
reasoneth, he does nothing else but conceive a sum total, from addition of
parcels; or conceive a remainder . . . These operations are not incident to
numbers only, but to all manner of things that can be added together, and
taken one out of another . . . the logicians teach the same in consequences
of words; adding together two names to make an affirmation, and to affirma
tions to make a syllogism; and many syllogisms to make a demonstration.”
5. gottfried wilhelm leibniz , characteristica universalis (1676).
6. n . p . padhy , artificial intelligence and intelligent systems 4 (2005,
2009).
7. dan w . patterson , introduction to artificial intelligence and expert
systems (1990).
8. george f . luger , artificial intelligence : structures and strategies
for complex problem solving (2001).
9. j . r . mcdonald , g . m . burt , j . s . zielinski , & s . d . j . mcarthur , intelli -
gent knowledge based system in electrical power engineering (1997).
10. stuart j . russell & peter norvig , artificial intelligence : a modern
approach (2002).
11. patterson , supra note 7.
12. steven l . tanimoto , elements of artificial intelligence : an intro -
duction using lisp (1987).
179
Note s to Chapte r 1 . 180
13. See, for example, Edwina L. Rissland, Artificial Intelligence and Law:
Stepping Stones to a Model of Legal Reasoning, 99 yale l . j . 1957, 1961 – 1964
(1990); alan tyree , expert systems in law 7 – 11 (1989).
14. robert m . glorioso & fernando c . colon osorio , engineering intelli -
gent systems : concepts and applications (1980).
15. padhy , supra note 6, at p. 13.
16. Nick Carbone, South Korea Rolls Out Robotic Prison Guards (Nov.
27, 2011), http://newsfeed.time.com (last visited Jan. 4, 2012); Alex Knapp,
South Korean Prison To Feature Robot Guards (Nov. 27, 2011), http://www
.forbes.com (last visited Feb. 29, 2012).
17. W. J. Hennigan, New Drone Has No Pilot Anywhere, So Who’s Ac-
countable?, los angeles times , Jan. 26, 2012. See also http://www.latimes
.com/business (last visited Feb. 29, 2012).
18. eugene charniak & drew mcdermott , introduction to artificial
intelligence (1985).
19. genesis 3:1 – 24.
20. richard e . bellman , an introduction to artificial intelligence :
can computers think ? (1978).
21. john haugeland , artificial intelligence : the very idea (1985).
22. charniak & mcdermott , supra note 18.
23. robert j . schalkoff , artificial intelligence : an engineering ap -
proach (1990).
24. raymond kurzweil , the age of intelligent machines (1990).
25. patrick henry winston , artificial intelligence (3d ed. 1992).
26. george f . luger & william a . stubblefield , artificial intelligence :
structures and strategies for complex problem solving (6th ed. 2008).
27. elaine rich & kevin knight , artificial intelligence (2d ed. 1991).
28. padhy , supra note 6, at p. 7.
29. Alan Turing, Computing Machinery and Intelligence, 59 mind 433,
433 – 460 (1950).
30. Donald Davidson, Turing’s Test, in modelling the mind 1 (K. A. Moh
yeldin Said, W. H. Newton-Smith, R. Viale, & K. V. Wilkes eds., 1990).
31. Robert M. French, Subcognition and the Limits of the Turing Test,
99 mind 53, 53 – 54 (1990).
32. masoud yazdani & ajit narayanan , artificial intelligence : human
effects (1985).
33. steven l . tanimoto , elements of artificial intelligence : an intro -
duction using lisp (1987).
34. john r . searle , minds , brains and science 28 – 41 (1984); John R.
Searle, Minds, Brains & Programs, 3 behavioral & brain sci . 417 (1980).
35. Roger C. Schank, What is ai , Anyway?, in the foundations of artifi -
cial intelligence 3, 4 – 6 (Derek Pertridge & Yorick Wilks eds., 1990, 2006).
36. douglas r . hofstadter , gödel , escher , bach : an eternal golden
braid 539 – 604 (1979, 1999).
37. See, for example, donald a . waterman , a guide to expert systems

(1986): “It wasn’t until the late 1970s that ai scientists began to realize
something quite important: The problem-solving power of a program comes
from the knowledge it possesses, not just from the formalisms and inference
schemes it employs. The conceptual breakthrough was made and can be
quite simply stated. To make a program intelligent, provide it with lots
of high-quality, specific knowledge about some problem area.” [original em
phasis]; donald michie & rory johnston , the creative computer (1984);
edward a . feigenbaum & pamela mccorduck , the fifth generation : artifi -
cial intelligence and japan ’ s computer challenge to the world (1983).
38. See, for example, the criticism of Winograd, supra note 3, at pp.
178 – 181.
39. Schank, supra note 35, pp. 9 – 12.
40. See phillip n . johnson -l aird , mental models 448 – 477 (1983); But see
also colin mcginn , the problem of consciousness : essays towards a reso -
lution 202, 209 – 213 (1991).
41. howard gardner , the mind ’ s new science : a history of the cognitive
revolution (1985); marvin minsky , the society of mind (1986); allen newell
& herbert a . simon , human problem solving (1972); Winograd, supra note 3,
at pp. 169 – 171.
42. max weber , economy and society : an outline of interpretive sociol -
ogy (1968); Winograd, supra note 3, at pp. 182 – 183.
43. Daniel C. Dennett, Evolution, Error, and Intentionality, in the foun -
dations of artificial intelligence 190, 190 – 211 (Derek Pertridge & Yorick
Wilks eds., 1990, 2006).
44. Lawrence B. Solum, Legal Personhood for Artificial Intelligences, 70
n . c . l . rev . 1231, 1262 (1992); owen j . flanagan , jr ., the science of the
mind 254 (2d ed. 1991); John Haugeland, Semantic Engines: An Introduction
to Mind Design, in mind design 1, 32 (John Haugeland ed., 1981).
45. monty newborn , deep blue (2002).
46. stephen baker , final jeopardy : man vs . machine and the quest to
know everything (2011).
47. vojislav kecman , learning and soft computing , support vector ma -
chines , neural networks and fuzzy logic models (2001).
48. For example, in November 2009, during the Supercomputing Con
ference in Portland, Oregon (SC09), IBM scientists and others announced
that they had succeeded in creating a new algorithm, named “Blue Matter,”
which possesses the thinking capabilities of a cat. Chris Capps, “Thinking”
Supercomputer Now Conscious as a Cat (Nov. 19, 2009), http://www
.unexplainable.net (last visited Feb. 29, 2012); International Conference for
High Performance Computing, Networking, Storage and Analysis, SC09,
http://sc09.supercomputing.org (last visited Feb. 29, 2012). This algorithm
collects information from very many units with parallel and distributed con
nections. The information is integrated, and creates a full image of sensory
information, perception, dynamic action and reaction, and cognition.

b . g . fitch et al ., blue matter : an application framework for molecular sim -
ulation on blue gene , IBM Research Report (2003). This platform simulates
brain capabilities, and eventually, it is supposed to simulate real thought pro
cesses. The final application of this algorithm contains not only analog and
digital circuits, metal or plastics, but also protein-based biologic surfaces.
49. See section 1.1.1.
50. terry winograd & fernando c . flores , understanding computers
and cognition : a new foundation for design (1986, 1987); Tom Athanasiou,
High-Tech Politics: The Case of Artificial Intelligence, 92 socialist rev . 7,
7 – 35 (1987); Winograd, supra note 3, at p. 181.
51. david levy , love and sex with robots : the evolution of human -
robot relationships (2007).
52. Yueh-Hsuan Weng, Chien-Hsun Chen, & Chuen-Tsai Sun, Toward the
Human-Robot Co-Existence Society: On Safety Intelligence for Next Genera-
tion Robots, 1 int . j . soc . robot . 267, 267 – 268 (2009).
53. International Robot Fair 2004 Organizing Office, World Robot Declara
tion (2004), http://www.prnewswire.co.uk (last visited Feb. 29, 2012).
54. Stefan Lovgren, A Robot in Every Home by 2020, South Korea Says,
national geographic news (2006), http://news.nationalgeographic.com (last
visited Feb. 29, 2012).
55. arthur c . clarke , 2001: a space odyssey (1968).
56. 2001: a space odyssey (Metro-Goldwyn-Mayer 1968).
57. terminator 2: judgment day (TriStar Pictures 1991).
58. the matrix (Warner Bros. Pictures 1999); the matrix reloaded
(Warner Bros. Pictures 2003); the matrix revolutions (Warner Bros. Pictures
2003).
59. i , robot (20th Century Fox 2004).
60. Yueh-Hsuan Weng, Chien-Hsun Chen, & Chuen-Tsai Sun, The Legal
Crisis of Next Generation Robots: On Safety Intelligence, Proceedings of the
11th International Conference on Artificial Intelligence and Law 205 – 209
(2007).
61. See, for example, Roboethics Roadmap Release 1.1, European Robot
ics Research Network (2006), http://www.roboethics.org (last visited Feb. 29,
2012).
62. isaac asimov , i , robot 40 (1950).
63. Jerry A. Fodor, Modules, Frames, Fridgeons, Sleeping Dogs, and the
Music of the Spheres, in the robot ’ s dilemma : the frame problem in artifi -
cial intelligence (Zenon W. Pylyshyn ed., 1987).
64. Isaac Asimov himself wrote in his introduction to The Rest of Robots
that “[t]here was just enough ambiguity in the Three Laws to provide the
conflicts and uncertainties required for new stories, and, to my great relief,
it seemed always to be possible to think up a new angle out of the sixty-one
words of the Three Laws.” issac asimov , the rest of robots 43 (1964).

66. Susan Leigh Anderson, Asimov’s “Three Laws of Robotics” and Ma-
chine Metaethics, 22 ai soc . 477 – 493 (2008).
67. Stefan Lovgren, Robot Codes of Ethics to Prevent Android Abuse,
Protect Humans, national geographic news (2007), http://news.national
geographic.com (last visited Feb. 29, 2012).
68. See, for example, gilbert ryle , the concept of mind 54 – 60 (1954);
richard b . brandt , ethical theory 389 (1959); john j . c . smart & bernard
williams , utilitarianism — for and against 18 – 20 (1973).
69. See, for example, Justine Miller, Criminal Law — An Agency for Social
Control, 43 yale l . j . 691 (1934); Jack P. Gibbs, A Very Short Step toward
a General Theory of Social Control am . b . found res . j . 607 (1985); K. W.
Lidstone, Social Control and the Criminal Law, 27 brit . j . criminology 31
(1987); Justice Ellis, Criminal Law as an Instrument of Social Control, 17 vic -
toria u . wellington l . rev . 319 (1987).
71. For the concept of turning disadvantages into advantages or advantag-
ing the disadvantages in industrial and private use, see section 1.1.3.
72. hans moravec , robot : mere machine to transcendent mind (1999).
73. david levy , robots unlimited : life in a virtual age (2006).
74. For wider aspect, see, for example, Mary Anne Warren, On the Moral
and Legal Status of Abortion, in ethics in practice (Hugh Lafollette ed.,
1997); immanuel kant , our duties to animals (1780).
75. David Lyons, Open Texture and the Possibility of Legal Interpretation,
18 law phil . 297, 297 – 309 (1999).
76. See, for example, exodus 22:1, 4, 9, 10; 23:4, 12.
77. exodus 21:28 – 32.
78. Jonathan M. E. Gabbai, Complexity and the Aerospace Industry: Un
derstanding Emergence by Relating Structure to Performance Using Multi-
Agent Systems (PhD Thesis, University of Manchester, 2005).
79. Craig W. Reynolds, Herds and Schools: A Distributed Behavioral
Model, 21 comput . graph . 25 – 34 (1987).
80. Wyatt S. Newman, Automatic Obstacle Avoidance at High Speeds via
Reflex Control, Proceedings of the 1989 IEEE International Conference on
Robotics and Automation 1104 (1989).
81. See, for example, Gabriel Hallevy, Unmanned Vehicles — Subordina-
tion to Criminal Law under the Modern Concept of Criminal Liability, 21 j . l .
inf . & sci . 311 (2011).
82. See, for example, State v. Stewart, 624 N.W.2d 585 (Minn.2001);
Wheatley v. Commonwealth, 26 Ky.L.Rep. 436, 81 S.W. 687 (1904); State v.
Follin, 263 Kan. 28, 947 P.2d 8 (1997).
83. Andrew G. Brooks & Ronald C. Arkin, Behavioral Overlays for Non-
Verbal Communication Expression on a Humanoid Robot, 22 auton . robots
55, 55 – 74 (2007).
84. lawrence lessig , code and other laws of cyberspace (1999).

85. See, for example, Steven J. Wolhandler, Voluntary Active Euthanasia
for the Terminally Ill and the Constitutional Right to Privacy, 69 cornell l .
rev . 363 (1984); Harold L. Hirsh & Richard E. Donovan, The Right to Die:
Medico-Legal Implications of In re Quinlan, 30 rutgers l . rev . 267 (1977);
Susan M. Allan, No Code Orders v. Resuscitation: The Decision to With-
hold Life-Prolonging Treatment from the Terminally Ill, 26 wayne l . rev . 139
(1980).
86. For the structure of the principle of legality in criminal law, see
gabriel hallevy , a modern treatise on the principle of legality in crimi -
nal law 5 – 8 (2010).
87. Id. at pp. 20 – 46.
88. Id. at pp. 67 – 78.
89. See, for example, in Transcript of Proceedings of Nuremberg Trials,
41 am . j . int ’ l l . 1 – 16 (1947).
90. hallevy , supra note 86, at pp. 97 – 118.
91. Id. at pp. 118 – 129.
92. Id. at pp. 135 – 137.
93. Id. at pp. 137 – 138.
94. Id. at pp. 138 – 141.
95. See, for example, sub-article 58(c)(1) of the Soviet Penal Code of 1926
as amended in 1950. This sub-article provided that mature relatives of the
first degree of convicted traitor are punished with five years of exile.
96. See supra note 89.
97. Scales v. United States, 367 U.S. 203, 81 S.Ct. 1469, 6 L.Ed.2d 782
(1961); Larsonneur, (1933) 24 Cr. App. R. 74, 97 J.P. 206, 149 L.T. 542;
andrew ashworth , principles of criminal law 106 – 107 (5th ed. 2006); An
derson v. State, 66 Okl.Cr. 291, 91 P.2d 794 (1939); State v. Asher, 50 Ark.
427, 8 S.W. 177 (1888); Peebles v. State, 101 Ga. 585, 28 S.E. 920 (1897);
Howard v. State, 73 Ga.App. 265, 36 S.E.2d 161 (1945); Childs v. State, 109
Nev. 1050, 864 P.2d 277 (1993).
98. See, for example, Smith v. State, 83 Ala. 26, 3 So. 551 (1888); People v.
Brubaker, 53 Cal.2d 37, 346 P.2d 8 (1959); State v. Barker, 128 W.Va. 744, 38
S.E.2d 346 (1946).
99. See, for example, Commonwealth v. Herd, 413 Mass. 834, 604 N.E.2d
1294 (1992); State v. Curry, 45 Ohio St.3d 109, 543 N.E.2d 1228 (1989); State
v. Barrett, 768 A.2d 929 (R.I.2001); State v. Lockhart, 208 W.Va. 622, 542
S.E.2d 443 (2000).
100. See, for example, Beason v. State, 96 Miss. 165, 50 So. 488 (1909);
State v. Nickelson, 45 La.Ann. 1172, 14 So. 134 (1893); Commonwealth v.
Mead, 92 Mass. 398 (1865); Willet v. Commonwealth, 76 Ky. 230 (1877);
Scott v. State, 71 Tex.Crim.R. 41, 158 S.W. 814 (1913); Price v. State, 50 Tex.
Crim.R. 71, 94 S.W. 901 (1906).
101. See, for example, Elk v. United States, 177 U.S. 529, 20 S.Ct. 729, 44
L.Ed. 874 (1900); State v. Bowen, 118 Kan. 31, 234 P. 46 (1925); Hughes v.
Commonwealth, 19 Ky.L.R. 497, 41 S.W. 294 (1897); People v. Cherry, 307
N.Y. 308, 121 N.E.2d 238 (1954); State v. Hooker, 17 Vt. 658 (1845); Common
wealth v. French, 531 Pa. 42, 611 A.2d 175 (1992).
102. gabriel hallevy , the matrix of derivative criminal liability 18 – 24
(2012).
103. Dugdale, (1853) 1 El. & Bl. 435, 118 E.R. 499, 500: “the mere intent
cannot constitute a misdemeanour when unaccompanied with any act”; Ex
parte Smith, 135 Mo. 223, 36 S.W. 628 (1896); Proctor v. State, 15 Okl.Cr.
338, 176 P. 771 (1918); State v. Labato, 7 N.J. 137, 80 A.2d 617 (1951); Lam
bert v. State, 374 P.2d 783 (Okla.Crim.App.1962); In re Leroy, 285 Md. 508,
403 A.2d 1226 (1979).
104. For the principle of legality in criminal law, see hallevy , supra note 86.
105. Id. at pp. 135 – 137.
106. Id. at pp. 49 – 80.
107. Id. at pp. 81 – 132.
108. See, for example, in the United States, Robinson v. California, 370
U.S. 660, 82 S.Ct. 1417, 8 L.Ed.2d 758 (1962).
109. glanville williams , criminal law : the general part sec. 11 (2nd ed.
1961).
110. andrew ashworth , principles of criminal law 157 – 158, 202 (5th ed.
2006).
111. G. R. Sullivan, Knowledge, Belief, and Culpability, in criminal law
theory — doctrines of the general part 207, 214 (Stephen Shute & A. P.
Simester eds., 2005).
112. See, for example, article 2.02(5) of The American Law Institute,
Model Penal Code — Official Draft and Explanatory Notes 22 (1962, 1985),
which provides, “When the law provides that negligence suffices to establish
an element of an offense, such element also is established if a person acts
purposely, knowingly or recklessly. When recklessness suffices to establish
an element, such element also is established if a person acts purposely or
knowingly. When acting knowingly suffices to establish an element, such
element also is established if a person acts purposely.”
113. William S. Laufer, Corporate Bodies and Guilty Minds, 43 emory l . j .
647 (1994); Kathleen F. Brickey, Corporate Criminal Accountability: A Brief
History and an Observation, 60 wash . u . l . q . 393 (1983).
114. william searle holdsworth , a history of english law 475 – 476
(1923).
115. William Searle Holdsworth, English Corporation Law in the 16th and
17th Centuries, 31 yale l . j . 382 (1922).
116. william robert scott , the constitution and finance of english ,
scottish and irish joint -s tock companies to 1720 462 (1912).
117. bishop carleton hunt , the development of the business corpora -
tion in england 1800 – 1 867 6 (1963).
118. See, for example, 6 Geo. I, c.18 (1719).

119. New York & G.L.R. Co. v. State, 50 N.J.L. 303, 13 A. 1 (1888); People
v. Clark, 8 N.Y.Cr. 169, 14 N.Y.S. 642 (1891); State v. Great Works Mill. & Mfg.
Co., 20 Me. 41, 37 Am.Dec.38 (1841); Commonwealth v. Proprietors of New
Bedford Bridge, 68 Mass. 339 (1854); Commonwealth v. New York Cent. &
H. River R. Co., 206 Mass. 417, 92 N.E. 766 (1910).
120. John C. Coffee, Jr., “No Soul to Damn: No Body to Kick”: An Unscan-
dalised Inquiry into the Problem of Corporate Punishment, 79 mich . l . rev .
386 (1981).
121. Langforth Bridge, (1635) Cro. Car. 365, 79 E.R. 919.
122. Clifton (Inhabitants), (1794) 5 T.R. 498, 101 E.R. 280; Great Brough
ton (Inhabitants), (1771) 5 Burr. 2700, 98 E.R. 418; Stratford-upon-Avon Cor
poration, (1811) 14 East 348, 104 E.R. 636; Liverpool (Mayor), (1802) 3 East
82, 102 E.R. 529; Saintiff, (1705) 6 Mod. 255, 87 E.R. 1002.
123. Severn and Wye Railway Co., (1819) 2 B. & Ald. 646, 106 E.R. 501;
Birmingham, &c., Railway Co., (1842) 3 Q. B. 223, 114 E.R. 492; New York
Cent. & H.R.R. v. United States, 212 U.S. 481, 29 S.Ct. 304, 53 L.Ed. 613
(1909); United States v. Thompson-Powell Drilling Co., 196 F.Supp. 571
(N.D.Tex.1961); United States v. Dye Construction Co., 510 F.2d 78 (10th
Cir.1975); United States v. Carter, 311 F.2d 934 (6th Cir.1963); State v. I. &
M. Amusements, Inc., 10 Ohio App.2d 153, 226 N.E.2d 567 (1966).
124. United States v. Alaska Packers’ Association, 1 Alaska 217 (1901).
125. United States v. John Kelso Co., 86 F. 304 (Cal.1898); Lennard’s Car
rying Co. Ltd. v. Asiatic Petroleum Co. Ltd., [1915] A.C. 705.
126. Director of Public Prosecutions v. Kent and Sussex Contractors Ltd.,
[1944] K.B. 146, [1944] 1 All E.R. 119; I.C.R. Haulage Ltd., [1944] K.B. 551,
[1944] 1 All E.R. 691; Seaboard Offshore Ltd. v. Secretary of State for Trans
port, [1994] 2 All E.R. 99, [1994] 1 W.L.R. 541, [1994] 1 Lloyd’s Rep. 593.
127. Granite Construction Co. v. Superior Court, 149 Cal.App.3d 465,
197 Cal.Rptr. 3 (1983); Commonwealth v. Fortner L.P. Gas Co., 610 S.W.2d
941 (Ky.App.1980); Commonwealth v. McIlwain School Bus Lines, Inc., 283
Pa.Super. 1, 423 A.2d 413 (1980); Gerhard O. W. Mueller, Mens Rea and the
Corporation — A Study of the Model Penal Code Position on Corporate Crim-
inal Liability, 19 u . pitt . l . rev . 21 (1957).
128. Hartson v. People, 125 Colo. 1, 240 P.2d 907 (1951); State v. Pincus,
41 N.J.Super. 454, 125 A.2d 420 (1956); People v. Sakow, 45 N.Y.2d 131, 408
N.Y.S.2d 27, 379 N.E.2d 1157 (1978).
2. ai Criminal Liability for I ntentional Offenses

1. The facts of this event are based on the overview in Yueh-Hsuan
Weng, Chien-Hsun Chen, & Chuen-Tsai Sun, Toward the Human-Robot Co-
Existence Society: On Safety Intelligence for Next Generation Robots, 1 int . j .
soc . robot 267, 273 (2009).

3. Fain v. Commonwealth, 78 Ky. 183, 39 Am.Rep. 213 (1879); Tift v.
State, 17 Ga.App. 663, 88 S.E. 41 (1916); People v. Decina, 2 N.Y.2d 133, 157
N.Y.S.2d 558, 138 N.E.2d 799 (1956); Mason v. State, 603 P.2d 1146 (Okl.
Crim.App.1979); State v. Burrell, 135 N.H. 715, 609 A.2d 751 (1992); Bonder
v. State, 752 A.2d 1169 (Del.2000).
4. The same is true with some other definitions, of which the most popu
lar is “willed muscular movement.” See herbert l . a . hart , punishment
and responsibility : essays in the philosophy of law 101 (1968); oliver w .
holmes , the common law 54 (1881, 1923); antony robin duff , criminal
attempts 239 – 263 (1996); john austin , the province of jurisprudence de -
termined (1832, 2000); gustav radbruch, der handlungsbegriff in seiner be-
deutung für das strafrechtssystem 75, 98 (1904); claus roxin, strafrecht —
allgemeiner teil i 239 – 255 (4 Auf. 2006); BGH 3, 287.
5. See, for example, Bolden v. State, 171 S.W.3d 785 (2005); United States
v. Meyers, 906 F. Supp. 1494 (1995); United States v. Quaintance, 471 F.
Supp.2d 1153 (2006).
6. Scott T. Noth, A Penny for Your Thoughts: Post-Mitchell Hate Crime
Laws Confirm a Mutating Effect upon Our First Amendment and the Govern-
ment’s Role in Our Lives, 10 regent u . l . rev . 167 (1998); henry holt , tele -
kinesis (2005); pamela rae heath , the pk zone : a cross -c ultural review of
psychokinesis ( pk ) (2003).
7. john austin , the province of jurisprudence determined (1832, 2000).
8. See, for example, People v. Heitzman, 9 Cal.4th 189, 37 Cal.Rptr.2d 236,
886 P.2d 1229 (1994); State v. Wilson, 267 Kan. 550, 987 P.2d 1060 (1999).
9. Rollin M. Perkins, Negative Acts in Criminal Law, 22 iowa l . rev . 659
(1937); Graham Hughes, Criminal Omissions, 67 yale l . j . 590 (1958); Lionel
H. Frankel, Criminal Omissions: A Legal Microcosm, 11 wayne l . rev . 367
(1965).
10. P. R. Glazebrook, Criminal Omissions: The Duty Requirement in Of-
fences Against the Person, 55 l . q . rev . 386 (1960); Andrew Ashworth, The
Scope of Criminal Liability for Omissions, 84 l . q . rev . 424, 441 (1989).
11. Lane v. Commonwealth, 956 S.W.2d 874 (Ky.1997); State v. Jackson,
137 Wash.2d 712, 976 P.2d 1229 (1999); Rachel S. Zahniser, Morally and Le-
gally: A Parent’s Duty to Prevent the Abuse of a Child as Defined by Lane v.
Commonwealth, 86 ky . l . j . 1209 (1998).
12. Mavji, [1987] 2 All E.R. 758, [1987] 1 W.L.R. 1388, [1986] S.T.C. 508,
Cr. App. Rep. 31, [1987] Crim. L.R. 39; Firth, (1990) 91 Cr. App. Rep. 217,
154 J.P. 576, [1990] Crim. L.R. 326.
13. See, for example, section 2.01(3) of The American Law Institute,
Model Penal Code — Official Draft and Explanatory Notes (1962, 1985).
14. gabriel hallevy , the matrix of derivative criminal liability 171 – 184
(2012).
15. See, for example, Pierson v. State, 956 P.2d 1119 (Wyo.1998).
16. See, for example, State v. Dubina, 164 Conn. 95, 318 A.2d 95 (1972);
State v. Bono, 128 N.J.Super. 254, 319 A.2d 762 (1974); State v. Fletcher, 322
N.C. 415, 368 S.E.2d 633 (1988).
17. S. Z. Feller, Les Délits de Mise en Danger, 40 rev . int . de droit pénal
179 (1969).
18. This is the results component of all homicide offenses. See sir edward
coke , institutions of the laws of england — third part 47 (6th ed. 1681,
1817, 2001): “Murder is when a man of sound memory, and of the age of dis
cretion, unlawfully killeth within any county of the realm any reasonable
creature in rerum natura under the king’s peace, with malice aforethought,
either expressed by the party or implied by law, [so as the party wounded,
or hurt, etc die of the wound or hurt, etc within a year and a day after the
same].”
19. Henderson v. Kibbe, 431 U.S. 145, 97 S.Ct. 1730, 52 L.Ed.2d 203
(1977); Commonwealth v. Green, 477 Pa. 170, 383 A.2d 877 (1978); State v.
Crocker, 431 A.2d 1323 (Me.1981); State v. Martin, 119 N.J. 2, 573 A.2d 1359
(1990).
20. See, for example, Wilson v. State, 24 S.W. 409 (Tex.Crim.App.1893);
Henderson v. State, 11 Ala.App. 37, 65 So. 721 (1914); Cox v. State, 305 Ark.
244, 808 S.W.2d 306 (1991); People v. Bailey, 451 Mich. 657, 549 N.W.2d 325
(1996).
21. Morton J. Horwitz, The Rise and Early Progressive Critique of Objec-
tive Causation, in the politics of law : a progressive critique 471 (David
Kairys ed., 3d ed. 1998); Benge, (1865) 4 F. & F. 504, 176 E.R. 665; Long
bottom, (1849) 3 Cox C. C. 439.
22. Jane Stapelton, Law, Causation and Common Sense, 8 oxford j . legal
stud . 111 (1988).
23. jerome hall , general principles of criminal law 70 – 77 (2d ed. 1960,
2005); david ormerod , smith & hogan criminal law 91 – 92 (11th ed. 2005);
G., [2003] U.K.H.L. 50, [2004] 1 A.C. 1034, [2003] 3 W.L.R. 1060, [2003] 4 All
E.R. 765, [2004] 1 Cr. App. Rep. 21, (2003) 167 J.P. 621, [2004] Crim. L. R. 369.
24. Sweet v. Parsley, [1970] A.C. 132, [1969] 1 All E.R. 347, [1969] 2 W.L.R.
470, 133 J.P. 188, 53 Cr. App. Rep. 221, 209 E.G. 703, [1969] E.G.D. 123.
25. See sections 1.3.2 and 2.1.
26. See, for example, supra note 16.
27. For the definition of circumstances, see section 2.1.2.
28. For the definition of results, see section 2.1.3.
29. This component of awareness functions also as the legal causal con
nection in mens rea offenses, but this function has no additional significance
in this context.
30. Specific intent is sometimes mistakenly referred to as intent in order
to differentiate it from general intent, which is normally used to express
mens rea.
32. sir gerald gordon , the criminal law of scotland 61 (1st ed. 1967);
Treacy v. Director of Public Prosecutions, [1971] A.C. 537, 559, [1971] 1 All
E.R. 110, [1971] 2 W.L.R. 112, 55 Cr. App. Rep. 113, 135 J.P. 112.
33. William G. Lycan, Introduction, in mind and cognition 3, 3 – 13 (Wil
liam G. Lycan ed., 1990).
34. See, for example, william james , the principles of psychology (1890).
35. See, for example, bernard baars , in the theatre of consciousness
(1997).
36. hermann von helmholtz , the facts of perception (1878).
37. United States v. Youts, 229 F.3d 1312 (10th Cir.2000); State v. Sargent,
156 Vt. 463, 594 A.2d 401 (1991); United States v. Spinney, 65 F.3d 231 (1st
Cir.1995); State v. Wyatt, 198 W.Va. 530, 482 S.E.2d 147 (1996); United States
v. Wert-Ruiz, 228 F.3d 250 (3rd Cir.2000); Rollin M. Perkins, “Knowledge” as
a Mens Rea Requirement, 29 hastings l . j . 953, 964 (1978).
38. United States v. Jewell, 532 F.2d 697 (9th Cir.1976); United States v.
Ladish Malting Co., 135 F.3d 484 (7th Cir.1998).
39. Paul Weiss, On the Impossibility of Artificial Intelligence, 44 rev .
metaphysics 335, 340 (1990).
40. See, for example, tim morris , computer vision and image process -
ing (2004); milan sonka , vaclav hlavac , & roger boyle , image processing ,
analysis , and machine vision (2008).
41. walter w . soroka , analog methods in computation and simulation
(1954).
42. See, for example, david manners & tsugio makimoto , living with the
chip (1995).
43. For robot guards, see section 1.1.1.
44. For the endlessness of the quest for machina sapiens, see section
1.1.2.
45. In re Winship, 397 U.S. 358, 90 S.Ct. 1068, 25 L.Ed.2d 368 (1970).
46. United States v. Heredia, 483 F.3d 913 (2006); United States v. Ramon-
Rodriguez, 492 F.3d 930 (2007); Saik, [2006] U.K.H.L. 18, [2007] 1 A.C. 18;
Da Silva, [2006] E.W.C.A. Crim. 1654 , [2006] 4 All E.R. 900, [2006] 2 Cr.
App. Rep. 517; Evans v. Bartlam, [1937] A.C. 473, 479, [1937] 2 All E.R.
646; G. R. Sullivan, Knowledge, Belief, and Culpability, in criminal law
theory — doctrines of the general part 207, 213 – 214 (Stephen Shute &
A. P. Simester eds., 2005).
47. State v. Pereira, 72 Conn. App. 545, 805 A.2d 787 (2002); Thompson v.
United States, 348 F.Supp.2d 398 (2005); Virgin Islands v. Joyce, 210 F. App.
208 (2006).
48. See, for example, United States v. Doe, 136 F.3d 631 (9th Cir.1998);
State v. Audette, 149 Vt. 218, 543 A.2d 1315 (1988); Ricketts v. State, 291
Md. 701, 436 A.2d 906 (1981); State v. Rocker, 52 Haw. 336, 475 P.2d 684
(1970); State v. Hobbs, 252 Iowa 432, 107 N.W.2d 238 (1961); State v. Daniels,
236 La. 998, 109 So.2d 896 (1958).
49. People v. Disimone, 251 Mich.App. 605, 650 N.W.2d 436 (2002);
Carter v. United States, 530 U.S. 255, 120 S.Ct. 2159, 147 L.Ed.2d 203 (2000);
State v. Neuzil, 589 N.W.2d 708 (Iowa 1999); People v. Henry, 239 Mich.App.
140, 607 N.W.2d 767 (1999); Frey v. United States, 708 So.2d 918 (Fla.1998);
United States v. Randolph, 93 F.3d 656 (9th Cir.1996); United States v.
Torres, 977 F.2d 321 (7th Cir.1992).
50. Schmidt v. United States, 133 F. 257 (9th Cir.1904); State v. Ehlers, 98
N.J.L. 263, 119 A. 15 (1922); United States v. Pomponio, 429 U.S. 10, 97 S.Ct.
22, 50 L.Ed.2d 12 (1976); State v. Gray, 221 Conn. 713, 607 A.2d 391 (1992);
State v. Mendoza, 709 A.2d 1030 (R.I.1998).
51. ludwig wittgenstein , philosophische untersuchungen §§629 – 660
(1953).
52. State v. Ayer, 136 N.H. 191, 612 A.2d 923 (1992); State v. Smith, 170
Wis.2d 701, 490 N.W.2d 40 (App.1992).
53. Studstill v. State, 7 Ga. 2 (1849); Glanville Williams, Oblique Inten-
tion, 46 camb . l . j . 417 (1987).
54. People v. Smith, 57 Cal. App. 4th 1470, 67 Cal. Rptr. 2d 604 (1997);
Wieland v. State, 101 Md. App. 1, 643 A.2d 446 (1994).
55. Stephen Shute, Knowledge and Belief in the Criminal Law, in crimi -
nal law theory — doctrines of the general part 182 – 187 (Stephen Shute
& A. P. Simester eds., 2005); antony kenny , will , freedom and power 42 – 43
(1975); john r . searle , the rediscovery of mind 62 (1992); antony kenny ,
what is faith ? 30 – 31 (1992); State v. VanTreese, 198 Iowa 984, 200 N.W.
570 (1924), but see Montgomery v. Commonwealth, 189 Ky. 306, 224 S.W.
878 (1920); State v. Murphy, 674 P.2d 1220 (Utah.1983); State v. Blakely, 399
N.W.2d 317 (S.D.1987).
56. See, for example, Ned Block, What Intuitions About Homunculi Don’t
Show, 3 behavioral & brain sci . 425 (1980); Bruce Bridgeman, Brains + Pro-
grams = Minds, 3 behavioral & brain sci . 427 (1980).
57. See, for example, feng -h siung hsu , behind deep blue : building the
computer that defeated the world chess champion (2002); david levy &
monty newborn , how computers play chess (1991).
59. For the endlessness of the quest for machina sapiens, see section
1.1.2.
61. Russian roulette is a game of chance in which participants place a
single round in a gun, spin the cylinder, place the muzzle against their own
head, and pull the trigger.
62. G., [2003] U.K.H.L. 50, [2003] 4 All E.R. 765, [2004] 1 Cr. App. Rep.
237, 167 J.P. 621, [2004] Crim. L.R. 369, [2004] 1 A.C. 1034; Victor Tadors,
Recklessness and the Duty to Take Care, in criminal law theory — doctrines
of the general part 227 (Stephen Shute & A. P. Simester eds., 2005); Gar
diner, [1994] Crim. L.R. 455.
63. vojislav kecman , learning and soft computing , support vector ma -

chines , neural networks and fuzzy logic models (2001).
64. K., [2001] U.K.H.L. 41, [2002] 1 A.C. 462; B. v. Director of Public Pros
ecutions, [2000] 2 A.C. 428, [2000] 1 All E.R. 833, [2000] 2 W.L.R. 452, [2000]
2 Cr. App. Rep. 65, [2000] Crim. L.R. 403.
65. Robert N. Shapiro, Of Robots, Persons, and the Protection of Religious
Beliefs, 56 s . cal . l . r . 1277, 1286 – 1290 (1983).
66. Nancy Sherman, The Place of the Emotions in Kantian Morality,
in identity , character , and morality 145, 149 – 1 62 (Owen Flanagan &
Amelie O. Rotry eds., 1990); Aaron Sloman, Motives, Mechanisms, and
Emotions, in the philosophy of artificial intelligence 231, 231 – 232
(Margaret A. Boden ed., 1990).
67. john finnis , natural law and natural rights 85 – 90 (1980).
n . c . l . rev . 1231, 1262 (1992).
69. owen j . flanagan , jr ., the science of the mind 224 – 241 (2d ed. 1991).
70. See, for example, david lyons , forms and limits of utilitarianism
(1965).
71. See Solum, supra note 68, at pp. 1258 – 1262.
73. edward o . wilson , sociobiology : the new synthesis 120 (1975).
74. roger j . sullivan , immanuel kant ’ s moral theory 68 (1989).
75. ray jackendoff , consciousness and the computational mind 275 – 327
(1987); colin mcginn , the problem of consciousness : essays towards a res -
olution 202 – 213 (1991).
76. daniel c . dennett , brainstorms 149 – 150 (1978).
78. daniel c . dennett , the intentional stance 327 – 328 (1987).
79. See, for example, People v. Marshall, 362 Mich. 170, 106 N.W.2d 842
(1961); State v. Gartland, 304 Mo. 87, 263 S.W. 165 (1924); State v. Etzweiler,
125 N.H. 57, 480 A.2d 870 (1984); People v. Kemp, 150 Cal.App.2d 654, 310
P.2d 680 (1957); State v. Hopkins, 147 Wash. 198, 265 P. 481 (1928); State v.
Foster, 202 Conn. 520, 522 A.2d 277 (1987); State v. Garza, 259 Kan. 826, 916
P.2d 9 (1996); Mendez v. State, 575 S.W.2d 36 (Tex.Crim.App.1979).
80. Francis Bowes Sayre, Criminal Responsibility for the Acts of Another,
43 harv . l . rev . 689, 689 – 690 (1930).
81. digesta , 9.4.2; ulpian , 18 ad ed.; olivia f . robinson , the criminal law
of ancient rome 15 – 16 (1995).
82. Y.BB. 32 – 33 Edw. I (R. S.), 318, 320 (1304); Seaman v. Browning,
(1589) 4 Leonard 123, 74 E.R. 771.
83. 13 Edw. I, St. I, c.2, art. 3, c.II, c.43 (1285). See also frederick pollock
& frederick william maitland , the history of english law before the time
of edward i 533 (rev. 2d ed. 1898); Oliver W. Holmes, Agency, 4 harv . l . rev .
345, 356 (1891).
84. Kingston v. Booth, (1685) Skinner 228, 90 E.R. 105.

85. Boson v. Sandford, (1690) 2 Salkeld 440, 91 E.R. 382: “The owners are
liable in respect of the freight, and as employing the master; for whoever em
ploys another is answerable for him, and undertakes for his care to all that
make use of him”; Turberwill v. Stamp, (1697) Skinner 681, 90 E.R. 303; Mid
dleton v. Fowler, (1699) 1 Salkeld 282, 91 E.R. 247; Jones v. Hart, (1699) 2
Salkeld 441, 91 E.R. 382; Hern v. Nichols, (1708) 1 Salkeld 289, 91 E.R. 256.
86. Sayre, supra note 80, at pp. 693 – 694; william paley , a treatise on
the law of principal and agent (2d ed. 1847); Huggins, (1730) 2 Strange
882, 93 E.R. 915; Holbrook, (1878) 4 Q.B.D. 42; Chisholm v. Doulton, (1889)
22 Q.B.D. 736; Hardcastle v. Bielby, [1892] 1 Q.B. 709.
87. Glanville Williams, Innocent Agency and Causation, 3 crim. l. f. 289
(1992); Peter Alldridge, The Doctrine of Innocent Agency, 2 crim. l. f. 45 (1990).
88. State v. Silva-Baltazar, 125 Wash.2d 472, 886 P.2d 138 (1994); glan -
ville williams , criminal law : the general part 395 (2d ed. 1961).
89. Maxey v. United States, 30 App. D.C. 63, 80 (App. D.C. 1907).
90. Johnson v. State, 142 Ala. 70, 71 (1904).
91. United States v. Bryan, 483 F.2d 88, 92 (3d Cir. 1973).
92. Commonwealth v. Hill, 11 Mass. 136 (1814); Michael, (1840) 2 Mood.
120, 169 E.R. 48.
93. Johnson v. State, 38 So. 182, 183 (Ala. 1904); People v. Monks, 24 P.2d
508, 511 (Cal. Dist. Ct. App. 1933).
94. United States v. Bryan, 483 F.2d 88, 92 (3d Cir. 1973); Boushea v.
United States, 173 F.2d 131, 134 (8th Cir. 1949); People v. Mutchler, 140 N.E.
820, 823 (Ill. 1923); State v. Runkles, 605 A.2d 111, 121 (Md. 1992); Parnell v.
State, 912 S.W.2d 422, 424 (Ark. 1996); State v. Thomas, 619 S.W.2d 513, 514
(Tenn. 1981).
95. Dusenbery v. Commonwealth, 263 S.E.2d 392 (Va. 1980).
96. United States v. Tobon-Builes, 706 F.2d 1092, 1101 (11th Cir. 1983);
United States v. Ruffin, 613 F.2d 408, 411 (2d Cir. 1979).
97. See Solum, supra note 68, at p. 1237.
98. The ai entity is used as an instrument and not as a participant, al
though it uses its features of processing information. See, for example,
George R. Cross & Cary G. Debessonet, An Artificial Intelligence Application
in the Law: CCLIPS, A Computer Program that Processes Legal Information,
1 high tech . l . j . 329 (1986).
99. Andrew J. Wu, From Video Games to Artificial Intelligence: Assign-
ing Copyright Ownership to Works Generated by Increasingly Sophisticated
Computer Programs, 25 aipla q . j . 131 (1997); Timothy L. Butler, Can a Com-
puter Be an Author — Copyright Aspects of Artificial Intelligence, 4 comm .
ent . l . s . 707 (1982).
100. nicola lacey & celia wells , reconstructing criminal law —
critical perspectives on crime and the criminal process 53 (2d ed. 1998).
101. See chapters 3 and 4.
102. People v. Monks, 133 Cal. App. 440, 446 (Cal. Dist. Ct. App. 1933).
103. See Solum, supra note 68, at pp. 1276 – 1279.
106. hallevy , supra note 14, at pp. 241 – 248.
107. digesta , 48.19.38.5; codex justinianus , 9.12.6; reinhard zimmermann ,
the law of obligations — roman foundations of the civilian tradition 197
(1996).
108. See, for example, United States v. Greer, 467 F.2d 1064 (7th Cir.1972);
People v. Cooper, 194 Ill.2d 419, 252 Ill.Dec. 458, 743 N.E.2d 32 (2000).
109. State v. Lucas, 55 Iowa 321, 7 N.W. 583 (1880); Roy v. United
States, 652 A.2d 1098 (D.C.App.1995); People v. Weiss, 256 App.Div. 162, 9
N.Y.S.2d 1 (1939); People v. Little, 41 Cal.App.2d 797, 107 P.2d 634 (1941).
110. State v. Linscott, 520 A.2d 1067 (Me.1987): “a rule allowing for a
murder conviction under a theory of accomplice liability based upon an
objective standard, despite the absence of evidence that the defendant pos
sessed the culpable subjective mental state that constitutes an element of
the crime of murder, does not represent a departure from prior Maine law”
(original emphasis).
111. People v. Prettyman, 14 Cal.4th 248, 58 Cal.Rptr.2d 827, 926 P.2d
1013 (1996); Chance v. State, 685 A.2d 351 (Del.1996); Ingram v. United
States, 592 A.2d 992 (D.C.App.1991); Richardson v. State, 697 N.E.2d 462
(Ind.1998); Mitchell v. State, 114 Nev. 1417, 971 P.2d 813 (1998); State v.
Carrasco, 122 N.M. 554, 928 P.2d 939 (1996); State v. Jackson, 137 Wash.2d
712, 976 P.2d 1229 (1999).
112. United States v. Powell, 929 F.2d 724 (D.C.Cir.1991).
113. State v. Kaiser, 260 Kan. 235, 918 P.2d 629 (1996); United States
v. Andrews, 75 F.3d 552 (9th Cir.1996); State v. Goodall, 407 A.2d 268
(Me.1979). Compare: People v. Kessler, 57 Ill.2d 493, 315 N.E.2d 29 (1974).
114. People v. Cabaltero, 31 Cal.App.2d 52, 87 P.2d 364 (1939); People v.
Michalow, 229 N.Y. 325, 128 N.E. 228 (1920).
115. Anderson, [1966] 2 Q.B. 110, [1966] 2 All E.R. 644, [1966] 2 W.L.R.
1195, 50 Cr. App. Rep. 216, 130 J.P. 318: “put the principle of law to be in
voked in this form: that where two persons embark on a joint enterprise,
each is liable for the acts done in pursuance of that joint enterprise, that that
includes liability for unusual consequences if they arise from the execution
of the agreed joint enterprise.”
116. English, [1999] A.C. 1, [1997] 4 All E.R. 545, [1997] 3 W.L.R. 959,
[1998] 1 Cr. App. Rep. 261, [1998] Crim. L.R. 48, 162 J.P. 1; Webb, [2006]
E.W.C.A. Crim. 2496, [2007] All E.R. (D) 406; O’Flaherty, [2004] E.W.C.A.
Crim. 526, [2004] 2 Cr. App. Rep. 315.
117. BGH 24, 213; BGH 26, 176; BGH 26, 244.
3. ai Criminal Liability for Negligence Offenses

1. For the role of ai expert systems in industry and research, see section
1.1.3.
2. See Gabriel Hallevy, The Criminal Liability of Artificial Intelligence
Entities — from Science Fiction to Legal Social Control, 4 akron intell .
prop . j . 171 (2010).
4. See section 2.1.
5. See reuven yaron , the laws of eshnunna 264 (2d ed. 1988).
6. collatio mosaicarum et romanarum legum , 1.6.1 – 4, 1.11.3 – 4; digesta ,
48.8.1.3, 48.19.5.2; ulpian , 7 de off. Proconsulis; pauli sententiae , 1 manual:
“magna neglegentia culpa est; magna culpa dolus est.”
7. henry de bracton , de legibus et consuetudinibus angliae 278 (1260;
G. E. Woodbine ed., S. E. Thorne trans., 1968 – 1977).
8. Hull, (1664) Kel. 40, 84 E.R. 1072, 1073.
9. Williamson, (1807) 3 Car. & P. 635, 172 E.R. 579.
10. Knight, (1828) 1 L.C.C. 168, 168 E.R. 1000; Grout, (1834) 6 Car. &
P. 629, 172 E.R. 1394; Dalloway, (1847) 2 Cox C.C. 273.
11. Finney, (1874) 12 Cox C.C. 625.
12. Bateman, [1925] All E.R. Rep. 45, 94 L.J.K.B. 791, 133 L.T. 730, 89 J.P.
162, 41 T.L.R. 557, 69 Sol. Jo. 622, 28 Cox. C.C. 33, 19 Cr. App. Rep. 8; Leach,
[1937] 1 All E.R. 319; Caldwell, [1982] A.C. 341, [1981] 1 All E.R. 961, [1981]
2 W.L.R. 509, 73 Cr. App. Rep. 13, 145 J.P. 211.
237, 167 J.P. 621, [2004] Crim. L.R. 369, [2004] 1 A.C. 1034.
14. jerome hall , general principles of criminal law 126 (2d ed. 1960,
2005).
15. Commonwealth v. Thompson, 6 Mass. 134, 6 Tyng 134 (1809); United
States v. Freeman, 25 Fed. Cas. 1208 (1827); Rice v. State, 8 Mo. 403 (1844);
United States v. Warner, 28 Fed. Cas. 404, 6 W.L.J. 255, 4 McLean 463 (1848);
Ann v. State, 30 Tenn. 159, 11 Hum. 159 (1850); State v. Schulz, 55 Ia. 628
(1881).
16. Lee v. State, 41 Tenn. 62, 1 Cold. 62 (1860); Chrystal v. Common
wealth, 72 Ky. 669, 9 Bush. 669 (1873).
17. Commonwealth v. Pierce, 138 Mass. 165 (1884); Abrams v. United
States, 250 U.S. 616, 63 L.Ed. 1173, 40 S.Ct. 17 (1919).
18. Commonwealth v. Walensky, 316 Mass. 383, 55 N.E.2d 902 (1944).
19. See, for example, People v. Haney, 30 N.Y.2d 328, 333 N.Y.S.2d 403,
284 N.E.2d 564 (1972); Leet v. State, 595 So.2d 959 (1991); Minor v. State,
326 Md. 436, 605 A.2d 138 (1992); United States v. Hanousek, 176 F.3d 1116
(9th Cir.1999).
20. See, for example, State v. Foster, 91 Wash.2d 466, 589 P.2d 789 (1979);
State v. Wilchinski, 242 Conn. 211, 700 A.2d 1 (1997); United States v.
Dominguez-Ochoa, 386 F.3d 639 (2004).
21. See, for example, Jerome Hall, Negligent Behaviour Should Be Ex-
cluded from Penal Liability, 63 colum . l . rev . 632 (1963); Robert P. Fine &
Gary M. Cohen, Is Criminal Negligence a Defensible Basis for Criminal Li-
ability?, 16 buff . l . r . 749 (1966).
23. For car accident statistics in the United States, see, for example,
http://www.cdc.gov (last visited Feb. 29, 2012).
26. For example, in France (article 121 – 3 of the French penal code).
28. In Hall v. Brooklands Auto Racing Club, [1932] All E.R. 208, [1933]
1 K.B. 205, 101 L.J.K.B. 679, 147 L.T. 404, 48 T.L.R. 546, the “reasonable
person” has been described this way: “The person concerned is sometimes
described as ‘the man in the street’, or ‘the man in the Clapham omnibus’, or,
as I recently read in an American author, ‘the man who takes the magazines
at home, and in the evening pushes the lawn mower in his shirt sleeves’.”
29. State v. Bunkley, 202 Conn. 629, 522 A.2d 795 (1987); State v. Evans,
134 N.H. 378, 594 A.2d 154 (1991).
30. People v. Decina, 2 N.Y.2d 133, 157 N.Y.S.2d 558, 138 N.E.2d 799
(1956); Government of the Virgin Islands v. Smith, 278 F.2d 169 (3rd
Cir.1960); People v. Howk, 56 Cal.2d 687, 16 Cal.Rptr. 370, 365 P.2d 426
(1961); State v. Torres, 495 N.W.2d 678 (1993).
33. See section 1.3.
34. See, for example, Rollins v. State, 2009 Ark. 484, 347 S.W.3d 20
(2009); People v. Larkins, 2010 Mich. App. Lexis 1891 (2010); Driver v. State,
2011 Tex. Crim. App. Lexis 4413 (2011).
36. See, for example, Peter Alldridge, The Doctrine of Innocent Agency,
2 crim . l . f . 45 (1990).
39. See sections 3.3.1 and 3.3.2.
40. See, for example, United States v. Robertson, 33 M.J. 832 (1991);
United States v. Buber, 62 M.J. 476 (2006).
4. ai Criminal Liability for Strict Liability Offenses

2. See, for example, W. J. Hennigan, New Drone Has No Pilot Anywhere,
So Who’s Accountable?, los angeles times , Jan. 26, 2012; see also
http://www.latimes.com (last visited Feb. 29, 2012).

4. See section 2.1.
5. Francis Bowes Sayre, Public Welfare Offenses, 33 colum . l . rev . 55, 56
(1933).
6. See, for example, Nutt, (1728) 1 Barn. K.B. 306, 94 E.R. 208; Dodd,
(1736) Sess. Cas. 135, 93 E.R. 136; Almon, (1770) 5 Burr. 2686, 98 E.R. 411;
Walter, (1799) 3 Esp. 21, 170 E.R. 524.
7. See, for example, 6 & 7 Vict. c.96.
8. Dixon, (1814) 3 M. & S. 11, 105 E.R. 516; Vantandillo, (1815) 4 M. & S.
73, 105 E.R. 762; Burnett, (1815) 4 M. & S. 272, 105 E.R. 835.
9. Woodrow, (1846) 15 M. & W. 404, 153 E.R. 907.
10. Stephens, [1866] 1 Q.B. 702; Fitzpatrick v. Kelly, [1873] 8 Q.B. 337;
Dyke v. Gower, [1892] 1 Q.B. 220; Blaker v. Tillstone, [1894] 1 Q.B. 345;
Spiers & Pond v. Bennett, [1896] 2 Q.B. 65; Hobbs v. Winchester Corporation,
[1910] 2 K.B. 471; Provincial Motor Cab Company Ltd. v. Dunning, [1909]
2 K.B. 599, 602.
11. W. G. Carson, Some Sociological Aspects of Strict Liability and the
Enforcement of Factory Legislation, 33 mod . l . rev . 396 (1970); W. G. Carson,
The Conventionalisation of Early Factory Crime, 7 int ’ l j . of sociology of
law 37 (1979).
12. austin turk , criminality and legal order (1969).
13. nicola lacey , celia wells , & oliver quick , reconstructing criminal
law 638 – 639 (3d ed. 2003, 2006).
14. Barnes v. State, 19 Conn. 398 (1849); Commonwealth v. Boynton, 84
Mass. 160, 2 Allen 160 (1861); Commonwealth v. Goodman, 97 Mass. 117
(1867); Farmer v. People, 77 Ill. 322 (1875); State v. Sasse, 6 S.D. 212, 60 N.W.
853 (1894); State v. Cain, 9 W. Va. 559 (1874); Redmond v. State, 36 Ark. 58
(1880); State v. Clottu, 33 Ind. 409 (1870); State v. Lawrence, 97 N.C. 492, 2
S.E. 367 (1887).
15. Myers v. State, 1 Conn. 502 (1816); Birney v. State, 8 Ohio Rep. 230
(1837); Miller v. State, 3 Ohio St. Rep. 475 (1854); Hunter v. State, 30 Tenn.
160, 1 Head 160 (1858); Stein v. State, 37 Ala. 123 (1861).
16. John R. Spencer & Antje Pedain, Approaches to Strict and Construc-
tive Liability in Continental Criminal Law, in appraising strict liability 237
(A. P. Simester ed., 2005).
17. Gammon (Hong Kong) Ltd. v. Attorney-General of Hong Kong, [1985]
1 A.C. 1, [1984] 2 All E.R. 503, [1984] 3 W.L.R. 437, 80 Cr. App. Rep. 194, 26
Build L.R. 159.
237, 167 J.P. 621, [2004] Crim. L.R. 369; Kumar, [2004] E.W.C.A. Crim. 3207,
[2005] 1 Cr. App. Rep. 566, [2005] Crim. L.R. 470; Matudi, [2004] E.W.C.A.
Crim. 697.
19. Salibaku v. France, (1998) E.H.R.R. 379.
20. 1950 European Human Rights Covenant, sec. 6(2) provides, “Everyone
charged with a criminal offence shall be presumed innocent until proved

guilty according to law.”
21. G., [2008] U.K.H.L. 37, [2009] A.C. 92; Barnfather v. Islington London
Borough Council, [2003] E.W.H.C. 418 (Admin), [2003] 1 W.L.R. 2318, [2003]
E.L.R. 263; G. R. Sullivan, Strict Liability for Criminal Offences in England
and Wales Following Incorporation into English Law of the European Con-
vention on Human Rights, in appraising strict liability 195 (A. P. Simester
ed., 2005).
22. Smith v. California, 361 U.S. 147, 80 S.Ct. 215, 4 L.Ed.2d 205 (1959);
Lambert v. California, 355 U.S. 225, 78 S.Ct. 240, 2 L.Ed.2d 228 (1957);
Texaco Inc. v. Short, 454 U.S. 516, 102 S.Ct. 781, 70 L.Ed.2d 738 (1982);
Carter v. United States, 530 U.S. 255, 120 S.Ct. 2159, 147 L.Ed.2d 203 (2000);
Alan C. Michaels, Imposing Constitutional Limits on Strict Liability: Lessons
from the American Experience, in appraising strict liability 218, 222 – 223
(A. P. Simester ed., 2005).
23. State v. Stepniewski, 105 Wis.2d 261, 314 N.W.2d 98 (1982); State
v. McDowell, 312 N.W.2d 301 (N.D. 1981); State v. Campbell, 536 P.2d 105
(Alaska 1975); Kimoktoak v. State, 584 P.2d 25 (Alaska 1978); Hentzner v.
State, 613 P.2d 821 (Alaska 1980); State v. Brown, 389 So.2d 48 (La.1980).
24. B. v. Director of Public Prosecutions, [2000] 2 A.C. 428, [2000] 1 All
E.R. 833, [2000] 2 W.L.R. 452, [2000] 2 Cr. App. Rep. 65, [2000] Crim. L.R.
403; Richards, [2004] E.W.C.A. Crim. 192.
25. For the negligence reasonability, see section 3.2.
27. In re Welfare of C.R.M., 611 N.W.2d 802 (Minn.2000); State v. Strong,
294 N.W.2d 319 (Minn.1980); Thompson v. State, 44 S.W.3d 171 (Tex.
App.2001); State v. Anderson, 141 Wash.2d 357, 5 P.3d 1247 (2000).
36. See, for example, Glanville Williams, Innocent Agency and Causation,
3 crim . l . f . 289 (1992).
38. Not in all legal systems is there a strict liability offense of homicide. In
such cases the relevant offense may be the particular traffic offense. See, for
example, Dana K. Cole, Expending Felony-Murder in Ohio: Felony-Murder or
Murder-Felony, 63 ohio st . l . j . 15 (2002).
41. See, for example, Peter Lynch, The Origins of Computer Weather Pre-
diction and Climate Modeling, 227 journal of computational physics 3431
(2008).
5. A pplicability of General Defenses to ai Criminal Liability

2. Nick Carbone, South Korea Rolls Out Robotic Prison Guards (Nov. 27,
2011), http://newsfeed.time.com (last visited Jan. 4, 2012); Alex Knapp,
South Korean Prison To Feature Robot Guards (Nov. 27, 2011), http://www
.forbes.com (last visited Feb. 29, 2012).
3. andrew ashworth , principles of criminal law 157 – 158, 202 (5th ed.
2006).
4. reuven yaron , the laws of eshnunna 265, 283 (2d ed. 1988).
5. Compare Kent Greenawalt, Distinguishing Justifications from Excuses,
49 law & contemp . probs . 89 (1986); Kent Greenawalt, The Perplexing Bor-
ders of Justification and Excuse, 84 colum . l . rev . 1897 (1984); george p .
fletcher , rethinking criminal law 759 – 817 (1978, 2000).
6. Compare Paul H. Robinson, A Theory of Justification: Societal Harm as
a Prerequisite for Criminal Liability, 23 u . c . l . a . l . rev . 266 (1975); Paul H.
Robinson, Testing Competing Theories of Justification, 76 n . c . l . rev . 1095
(1998); George P. Fletcher, The Nature of Justification, in action and value
in criminal law 175 (Stephen Shute, John Gardner, & Jeremy Horder eds.,
2003).
7. See section 5.1.
8. rudolph sohm , the institutes of roman law 219 (3d ed. 1907).
9. See, for example, Minn. Stat. §9913 (1927); Mont. Rev. Code §10729
(1935); N.Y. Penal Code §816 (1935); Okla. Stat. §152 (1937); Utah Rev. Stat.
103-i-40 (1933).
10. State v. George, 20 Del. 57, 54 A. 745 (1902); Heilman v. Common
wealth, 84 Ky. 457, 1 S.W. 731 (1886); State v. Aaron, 4 N.J.L. 269 (1818).
11. State v. Dillon, 93 Idaho 698, 471 P.2d 553 (1970); State v. Jackson, 346
Mo. 474, 142 S.W.2d 45 (1940).
12. See Godfrey v. State, 31 Ala. 323 (1858); Martin v. State, 90 Ala. 602,
8 So. 858 (1891); State v. J.P.S., 135 Wash.2d 34, 954 P.2d 894 (1998); Beason
v. State, 96 Miss. 165, 50 So. 488 (1909); State v. Nickelson, 45 La.Ann. 1172,
14 So. 134 (1893); Commonwealth v. Mead, 92 Mass. 398 (1865); Willet v.
Commonwealth, 76 Ky. 230 (1877); Scott v. State, 71 Tex.Crim.R. 41, 158
S.W. 814 (1913); Price v. State, 50 Tex.Crim.R. 71, 94 S.W. 901 (1906).
13. Adams v. State, 8 Md.App. 684, 262 A.2d 69 (1970): “the most modern
definition of the test is simply that the surrounding circumstances must dem
onstrate, beyond a reasonable doubt, that the individual knew what he was
doing and that it was wrong.”
14. A. W. G. Kean, The History of the Criminal Liability of Children, 53
l . q . rev . 364 (1937).
15. Andrew Walkover, The Infancy Defense in the New Juvenile Court, 31
u . c . l . a . l . rev .
503 (1984); Keith Foren, Casenote: In Re Tyvonne M. Revis-
ited: The Criminal Infancy Defense in Connecticut, 18 q . l . rev . 733 (1999).
16. Frederick J. Ludwig, Rationale of Responsibility for Young Offenders, 29
neb . l . rev . 521 (1950); In re Tyvonne, 211 Conn. 151, 558 A.2d 661 (1989).
19. In Bratty v. Attorney-General for Northern Ireland, [1963] A.C. 386,
409, [1961] 3 All E.R. 523, [1961] 3 W.L.R. 965, 46 Cr. App. Rep 1, Lord Den
ning noted, “The requirement that it should be a voluntary act is essential,
not only in a murder case, but also in every criminal case. No act is punish
able if it is done involuntarily.” State v. Mishne, 427 A.2d 450 (Me.1981);
State v. Case, 672 A.2d 586 (Me.1996).
20. See, for example, People v. Newton, 8 Cal.App.3d 359, 87 Cal.Rptr.
394 (1970).
21. Kenneth L. Campbell, Psychological Blow Automatism: A Narrow De-
fence, 23 crim . l . q . 342 (1981); Winifred H. Holland, Automatism and Crim-
inal Responsibility, 25 crim . l . q . 95 (1982).
22. People v. Higgins, 5 N.Y.2d 607, 186 N.Y.S.2d 623, 159 N.E.2d 179
(1959); State v. Welsh, 8 Wash.App. 719, 508 P.2d 1041 (1973).
23. Reed v. State, 693 N.E.2d 988 (Ind.App.1998).
24. Quick, [1973] Q.B. 910, [1973] 3 All E.R. 347, [1973] 3 W.L.R. 26, 57
Cr. App. Rep. 722, 137 J.P. 763; C, [2007] E.W.C.A. Crim. 1862, [2007] All
E.R. (D) 91.
25. Fain v. Commonwealth, 78 Ky. 183 (1879); Bradley v. State, 102 Tex.
Crim.R. 41, 277 S.W. 147 (1926); Norval Morris, Somnambulistic Homicide:
Ghosts, Spiders, and North Koreans, 5 res judicatae 29 (1951).
26. McClain v. State, 678 N.E.2d 104 (Ind.1997).
27. People v. Newton, 8 Cal.App.3d 359, 87 Cal.Rptr. 394 (1970); Read v.
People, 119 Colo. 506, 205 P.2d 233 (1949); Carter v. State, 376 P.2d 351 (Okl.
Crim.App.1962).
28. People v. Wilson, 66 Cal.2d 749, 59 Cal.Rptr. 156, 427 P.2d 820 (1967);
People v. Lisnow, 88 Cal.App.3d Supp. 21, 151 Cal.Rptr. 621 (1978); Law
rence Taylor & Katharina Dalton, Premenstrual Syndrome: A New Criminal
Defense?, 19 cal . w . l . rev . 269 (1983); Michael J. Davidson, Feminine Hor-
monal Defenses: Premenstrual Syndrome and Postpartum Psychosis, 5 army
lawyer (2000).
29. Government of the Virgin Islands v. Smith, 278 F.2d 169 (3rd
Cir.1960); People v. Freeman, 61 Cal.App.2d 110, 142 P.2d 435 (1943); State
v. Hinkle, 200 W.Va. 280, 489 S.E.2d 257 (1996).
30. State v. Gish, 17 Idaho 341, 393 P.2d 342 (1964); Evans v. State, 322
Md. 24, 585 A.2d 204 (1991); State v. Jenner, 451 N.W.2d 710 (S.D.1990);
Lester v. State, 212 Tenn. 338, 370 S.W.2d 405 (1963); Polston v. State, 685
P.2d 1 (Wyo.1984).
31. Richard Delgado, Ascription of Criminal States of Mind: Toward a De-

fense Theory for the Coercively Persuaded (“Brainwashed”) Defendant, 63
minn . l . rev . 1 (1978); Joshua Dressler, Professor Delgado’s “Brainwashing”
Defense: Courting a Determinist Legal System, 63 minn . l . rev . 335 (1978).
32. francis antony whitlock , criminal responsibility and mental ill -
ness 119 – 120 (1963).
33. RG 60, 29; RG 73, 177; VRS 23, 212; VRS 46, 440; VRS 61, 339; VRS
64, 189; DAR 1985, 387; BGH 2, 14; BGH 17, 259; BGH 21, 381.
34. karl menninger , martin mayman , & paul pruyser , the vital balance
420 – 489 (1963); George Mora, Historical and Theoretical Trends in Psychia-
try, in comprehensive textbook of psychiatry 1, 8 – 19 (Alfred M. Freedman,
Harold Kaplan, & Benjamin J. Sadock eds., 2d ed. 1975).
35. michael moore , law and psychiatry : rethinking the relationship
64 – 65 (1984); Anthony Platt & Bernard L. Diamond, The Origins of the
“Right and Wrong” Test of Criminal Responsibility and Its Subsequent De-
velopment in the United States: An Historical Survey, 54 cal . l . rev . 1227
(1966).
36. sander l . gilman , seeing the insane (1982); john biggs , the guilty
mind 26 (1955).
37. walter bromberg , from shaman to psychotherapist : a history of
the treatment of mental illness 63 (1975); george rosen , madness in soci -
ety : chapters in the historical sociology of mental illness 33, 82 (1969);
edward norbeck , religion in primitive society 215 (1961).
38. james cowles prichard , a treatise on insanity and other disorders
affecting the mind (1835); arthur e . fink , causes of crime : biological the -
ories in the united states , 1800 – 1 915 48 – 76 (1938); Janet A. Tighe, Fran-
cis Wharton and the Nineteenth Century Insanity Defense: The Origins of a
Reform Tradition, 27 am . j . legal hist . 223 (1983).
39. Peter McCandless, Liberty and Lunacy: The Victorians and Wrongful
Confinement, in madhouses , mad -d octors , and madmen : the social history
of psychiatry in the victorian era 339, 354 (Scull ed., 1981); vieda skul -
tans , english madness : ideas on insanity , 1580 – 1 890 69 – 97 (1979); michel
foucault , madness and civilization 24 (1965).
40. Seymour L. Halleck, The Historical and Ethical Antecedents of Psy-
chiatric Criminology, in psychiatric aspects of criminology 8 (Halleck &
Bromberg eds., 1968); franz alexander & hugo staub , the criminal , the
judge , and the public 24 – 25 (1931); franz alexander , our age of unreason :
a study of the irrational forces in social life (rev. ed. 1971).
41. Homer D. Crotty, The History of Insanity as a Defence to Crime in En-
glish Common Law, 12 cal . l . rev . 105, 107 – 108 (1924).
42. M’Naghten, (1843) 10 Cl. & Fin. 200, 8 E.R. 718.
43. Oxford, (1840) 9 Car. & P. 525, 173 E.R. 941.
45. United States v. Freeman, 357 F.2d 606 (2nd Cir.1966); United States
v. Currens, 290 F.2d 751 (3rd Cir.1961); United States v. Chandler, 393 F.2d
920 (4th Cir.1968); Blake v. United States, 407 F.2d 908 (5th Cir.1969); United
States v. Smith, 404 F.2d 720 (6th Cir.1968); United States v. Shapiro, 383
F.2d 680 (7th Cir.1967); Pope v. United States, 372 F.2d 710 (8th Cir.1970).
46. Commonwealth v. Herd, 413 Mass. 834, 604 N.E.2d 1294 (1992); State
v. Curry, 45 Ohio St.3d 109, 543 N.E.2d 1228 (1989); State v. Barrett, 768
A.2d 929 (R.I.2001); State v. Lockhart, 208 W.Va. 622, 542 S.E.2d 443 (2000).
See also 18 U.S.C.A. §17.
47. The American Law Institute, Model Penal Code — Official Draft and
Explanatory Notes 61 – 62 (1962, 1985): “(1) A person is not responsible for
criminal conduct if at the time of such conduct as a result of mental disease
or defect he lacks substantial capacity either to appreciate the criminality
[wrongfulness] of his conduct or to conform his conduct to the requirements
of law; (2) As used in this Article, the terms ‘mental disease or defect’ do not
include an abnormality manifested only by repeated criminal or otherwise
antisocial conduct.”
48. State v. Elsea, 251 S.W.2d 650 (Mo.1952); State v. Johnson, 233 Wis.
668, 290 N.W. 159 (1940); State v. Hadley, 65 Utah 109, 234 P. 940 (1925);
henry weihofen , mental disorder as a criminal defense 119 (1954);
K. W. M. Fulford, Value, Action, Mental Illness, and the Law, in action and
value in criminal law 279 (Stephen Shute, John Gardner, & Jeremy Horder
eds., 2003).
50. People v. Sommers, 200 P.3d 1089 (2008); McNeil v. United States, 933
A.2d 354 (2007); Rangel v. State, 2009 Tex.App. 1555 (2009); Commonwealth
v. Shumway, 72 Va.Cir. 481 (2007).
51. R. U. Singh, History of the Defence of Drunkenness in English Crimi-
nal Law, 49 law q . rev . 528 (1933).
52. theodori liber poenitentialis , III, 13 (668 – 690).
53. Francis Bowes Sayre, Mens Rea, 45 harv . l . rev . 974, 1014 – 1015
(1932).
54. william oldnall russell , a treatise on crimes and misdemeanors 8
(1843, 1964).
55. Marshall, (1830) 1 Lewin 76, 168 E.R. 965.
56. Pearson, (1835) 2 Lewin 144, 168 E.R. 1108; Thomas, (1837) 7 Car. & P.
817, 173 E.R. 356: “Drunkenness may be taken into consideration in cases
where what the law deems sufficient provocation has been given, because
the question is, in such cases, whether the fatal act is to be attributed to the
passion of anger excited by the previous provocation, and that passion is
more easily excitable in a person when in a state of intoxication than when
he is sober.”
57. Meakin, (1836) 7 Car. & P. 297, 173 E.R. 131; Meade, [1909] 1 K.B. 895;
Pigman v. State, 14 Ohio 555 (1846); People v. Harris, 29 Cal. 678 (1866);
People v. Townsend, 214 Mich. 267, 183 N.W. 177 (1921).
58. Derrick Augustus Carter, Bifurcations of Consciousness: The Elimi-

nation of the Self-Induced Intoxication Excuse, 64 mo . l . rev . 383 (1999);
Jerome Hall, Intoxication and Criminal Responsibility, 57 harv . l . rev . 1045
(1944); Monrad G. Paulsen, Intoxication as a Defense to Crime, 1961 u . ill .
l . f . 1 (1961).
59. State v. Cameron, 104 N.J. 42, 514 A.2d 1302 (1986); State v. Smith,
260 Or. 349, 490 P.2d 1262 (1971); People v. Leonardi, 143 N.Y. 360, 38 N.E.
372 (1894); Tate v. Commonwealth, 258 Ky. 685, 80 S.W.2d 817 (1935); Rob
erts v. People, 19 Mich. 401 (1870); People v. Kirst, 168 N.Y. 19, 60 N.E. 1057
(1901); State v. Robinson, 20 W.Va. 713, 43 Am.Rep. 799 (1882).
60. Addison M. Bowman, Narcotic Addiction and Criminal Responsibility
under Durham, 53 geo . l . j . 1017 (1965); Herbert Fingarette, Addiction and
Criminal Responsibility, 84 yale l . j . 413 (1975); Lionel H. Frankel, Narcotic
Addiction, Criminal Responsibility and Civil Commitment, 1966 utah l . rev .
581 (1966); Peter Barton Hutt & Richard A. Merrill, Criminal Responsibility
and the Right to Treatment for Intoxication and Alcoholism, 57 geo . l . j . 835
(1969).
61. Powell v. Texas, 392 U.S. 514, 88 S.Ct. 2145, 20 L.Ed.2d 1254 (1968);
United States v. Moore, 486 F.2d 1139 (D.C.Cir.1973); State v. Herro, 120
Ariz. 604, 587 P.2d 1181 (1978); State v. Smith, 219 N.W.2d 655 (Iowa 1974);
People v. Davis, 33 N.Y.2d 221, 351 N.Y.S.2d 663, 306 N.E.2d 787 (1973).
64. William G. Lycan, Introduction, in mind and cognition 3 – 13
(William G. Lycan ed., 1990).
65. Edwin R. Keedy, Ignorance and Mistake in the Criminal Law, 22
harv . l . rev . 75, 78 (1909); Levett, (1638) Cro. Car. 538.
66. State v. Silveira, 198 Conn. 454, 503 A.2d 599 (1986); State v. Molin, 288
N.W.2d 232 (Minn.1979); State v. Sexton, 160 N.J. 93, 733 A.2d 1125 (1999).
67. State v. Sawyer, 95 Conn. 34, 110 A. 461 (1920); State v. Cude, 14 Utah
2d 287, 383 P.2d 399 (1963); Ratzlaf v. United States, 510 U.S. 135, 114 S.Ct.
655, 126 L.Ed.2d 615 (1994); Cheek v. United States, 498 U.S. 192, 111 S.Ct.
604, 112 L.Ed.2d 617 (1991); Richard H. S. Tur, Subjectivism and Objectiv-
ism: Towards Synthesis, in action and value in criminal law 213 (Stephen
Shute, John Gardner, & Jeremy Horder eds., 2003).
68. United States v. Lampkins, 4 U.S.C.M.A. 31, 15 C.M.R. 31 (1954).
69. People v. Vogel, 46 Cal.2d 798, 299 P.2d 850 (1956); Long v. State, 44
Del. 262, 65 A.2d 489 (1949).
70. Fernand N. Dutile & Harold F. Moore, Mistake and Impossibility:
Arranging Marriage Between Two Difficult Partners, 74 nw . u . l . rev . 166
(1980).
71. Douglas Husak & Andrew von Hirsch, Culpability and Mistake of Law,
in action and value in criminal law 157, 161 – 167 (Stephen Shute, John
Gardner, & Jeremy Horder eds., 2003).
72. digesta , 22.6.9: “juris quidam ignorantiam cuique nocere, facti vero
ignorantiam non nocere.”
73. See, for example, Brett v. Rigden, (1568) 1 Plowd. 340, 75 E.R. 516;
Mildmay, (1584) 1 Co. Rep. 175a, 76 E.R. 379; Manser, (1584) 2 Co. Rep. 3,
76 E.R. 392; Vaux, (1613) 1 Blustrode 197, 80 E.R. 885; Bailey, (1818) Russ. &
Ry. 341, 168 E.R. 835; Esop, (1836) 7 Car. & P. 456, 173 E.R. 203; Crawshaw,
(1860) Bell. 303, 169 E.R. 1271; Schuster v. State, 48 Ala. 199 (1872).
74. Forbes, (1835) 7 Car. & P. 224, 173 E.R. 99; Parish, (1837) 8 Car. & P. 94,
173 E.R. 413; Allday, (1837) 8 Car. & P. 136, 173 E.R. 431; Dotson v. State, 6
Cold. 545 (1869); Cutter v. State, 36 N.J.L. 125 (1873); Squire v. State, 46 Ind.
459 (1874).
75. State v. Goodenow, 65 Me. 30 (1876); State v. Whitoomb, 52 Iowa 85,
2 N.W. 970 (1879).
76. Lutwin v. State, 97 N.J.L. 67, 117 A. 164 (1922); State v. Whitman,
116 Fla. 196, 156 So. 705 (1934); United States v. Mancuso, 139 F.2d 90 (3rd
Cir.1943); State v. Chicago, M. & St.P.R. Co., 130 Minn. 144, 153 N.W. 320
(1915); Coal & C.R. v. Conley, 67 W.Va. 129, 67 S.E. 613 (1910); State v. Strig
gles, 202 Iowa 1318, 210 N.W. 137 (1926); United States v. Albertini, 830 F.2d
985 (9th Cir.1987).
77. State v. Sheedy, 125 N.H. 108, 480 A.2d 887 (1984); People v. Fergu
son, 134 Cal.App. 41, 24 P.2d 965 (1933); Andrew Ashworth, Testing Fidelity
to Legal Values: Official Involvement and Criminal Justice , 63 mod . l . rev .
663 (2000); Glanville Williams, The Draft Code and Reliance upon Official
Statements, 9 legal stud . 177 (1989).
78. Rollin M. Perkins, Ignorance and Mistake in Criminal Law, 88 u . pa . l .
rev . 35 (1940).
83. Chas E. George, Limitation of Police Powers, 12 law . & banker &
s . bench & b . rev . 740 (1919); Kam C. Wong, Police Powers and Control in
the People’s Republic of China: The History of Shoushen, 10 colum . j . asian
l . 367 (1996); John S. Baker, Jr., State Police Powers and the Federalization of
Local Crime, 72 temp . l . rev . 673 (1999).
84. Dolores A. Donovan & Stephanie M. Wildman, Is the Reasonable Man
Obsolete? A Critical Perspective on Self-Defense and Provocation, 14 loy .
l . a . l . rev . 435, 441 (1981); Joshua Dressler, Rethinking Heat of Passion: A
Defense in Search of a Rationale, 73 j . crim . l . & criminology 421, 444 – 450
(1982); Kent Greenawalt, The Perplexing Borders of Justification and Excuse,
84 colum . l . rev . 1897, 1898, 1915 – 1919 (1984).
85. State v. Brosnan, 221 Conn. 788, 608 A.2d 49 (1992); State v. Galla
gher, 191 Conn. 433, 465 A.2d 323 (1983); State v. Nelson, 329 N.W.2d 643
(Iowa 1983); State v. Farley, 225 Kan. 127, 587 P.2d 337 (1978).
86. Commonwealth v. Monico, 373 Mass. 298, 366 N.E.2d 1241 (1977);
Commonwealth v. Johnson, 412 Mass. 368, 589 N.E.2d 311 (1992); Duck
ett v. State, 966 P.2d 941 (Wyo.1998); People v. Young, 11 N.Y.2d 274, 229
N.Y.S.2d 1, 183 N.E.2d 319 (1962); Batson v. State, 113 Nev. 669, 941 P.2d
478 (1997); State v. Wenger, 58 Ohio St.2d 336, 390 N.E.2d 801 (1979); Moore
v. State, 25 Okl.Crim. 118, 218 P. 1102 (1923).
87. Williams v. State, 70 Ga.App. 10, 27 S.E.2d 109 (1943); State v.
Totman, 80 Mo.App. 125 (1899).
88. Lawson, [1986] V.R. 515; Daniel v. State, 187 Ga. 411, 1 S.E.2d 6
(1939).
89. John Barker Waite, The Law of Arrest, 24 tex . l . rev . 279 (1946).
90. People v. Williams, 56 Ill.App.2d 159, 205 N.E.2d 749 (1965); People
v. Minifie, 13 Cal.4th 1055, 56 Cal.Rptr.2d 133, 920 P.2d 1337 (1996); State v.
Coffin, 128 N.M. 192, 991 P.2d 477 (1999).
91. State Philbrick, 402 A.2d 59 (Me.1979); State v. Havican, 213 Conn.
593, 569 A.2d 1089 (1990); State v. Harris, 222 N.W.2d 462 (Iowa 1974);
Judith Fabricant, Homicide in Response to a Threat of Rape: A Theoreti-
cal Examination of the Rule of Justification, 11 golden gate u . l . rev . 945
(1981).
92. Celia Wells, Battered Woman Syndrome and Defences to Homicide:
Where Now?, 14 legal stud . 266 (1994); Aileen McColgan, In Defence of Bat-
tered Women who Kill, 13 oxford j . legal stud . 508 (1993); Joshua Dressler,
Battered Women Who Kill Their Sleeping Tormenters: Reflections on Main-
taining Respect for Human Life while Killing Moral Monsters, in criminal
law theory — doctrines of the general part 259 (Stephen Shute & A. P.
Simester eds., 2005).
93. State v. Moore, 158 N.J. 292, 729 A.2d 1021 (1999); State v. Robinson,
132 Ohio App.3d 830, 726 N.E.2d 581 (1999).
n . c . l . rev . 1231, 1255 – 1258 (1992).
95. See section 1.2.1; isaac asimov , i , robot 40 (1950).
96. See, for example, United States v. Allegheny Bottling Company, 695
F.Supp. 856 (1988); John C. Coffee, Jr., “No Soul to Damn: No Body to Kick”:
An Unscandalised Inquiry into the Problem of Corporate Punishment, 79
mich . l . rev . 386 (1981).
97. judith jarvis thomson , rights , restitution and risk : essays in moral
theory 33 – 48 (1986); Sanford Kadish, Respect for Life and Regard for Rights
in the Criminal Law, 64 cal . l . rev . 871 (1976); Patrick Montague, Self-
Defense and Choosing Between Lives, 40 phil . stud . 207 (1981); Cheyney C.
Ryan, Self-Defense, Pacificism, and the Possibility of Killing, 93 ethics 508
(1983).

102. See, for example, United States v. Holmes, 26 F. Cas. 360, 1 Wall. Jr. 1
(1842); Dudley and Stephens, [1884] 14 Q.B. D. 273.
103. W. H. Hitchler, Necessity as a Defence in Criminal Cases, 33 dick .
l . rev . 138 (1929).
104. Edward B. Arnolds & Norman F. Garland, The Defense of Necessity in
Criminal Law: The Right to Choose the Lesser Evil, 65 j . crim . l . & criminol -
ogy 289 (1974); Lawrence P. Tiffany & Carl A. Anderson, Legislating the Ne-
cessity Defense in Criminal Law, 52 denv . l . j . 839 (1975); Rollin M. Perkins,
Impelled Perpetration Restated, 33 hastings l . j . 403 (1981).
105. Long v. Commonwealth, 23 Va.App. 537, 478 S.E.2d 324 (1996);
State v. Crocker, 506 A.2d 209 (Me.1986); Humphrey v. Commonwealth,
37 Va.App. 36, 553 S.E.2d 546 (2001); United States v. Oakland Cannabis
Buyers’ Cooperative, 532 U.S. 483, 121 S.Ct. 1711, 149 L.Ed.2d 722 (2001);
United States v. Kabat, 797 F.2d 580 (8th Cir.1986); McMillan v. City of Jack
son, 701 So.2d 1105 (Miss.1997).
106. benjamin thorpe , ancient laws and institutes of england 47 – 49
(1840, 2004); Reniger v. Fogossa, (1551) 1 Plowd. 1, 75 E.R. 1, 18; Mouse,
(1608) 12 Co. Rep. 63, 77 E.R. 1341; michael dalton , the countrey justice
ch. 150 (1618, 2003).
107. United States v. Randall, 104 Wash.D.C.Rep. 2249 (D.C.Super.1976);
State v. Hastings, 118 Idaho 854, 801 P.2d 563 (1990); People v. Whipple, 100
Cal.App. 261, 279 P. 1008 (1929); United States v. Paolello, 951 F.2d 537 (3rd
Cir.1991).
108. Commonwealth v. Weaver, 400 Mass. 612, 511 N.E.2d 545 (1987);
Nelson v. State, 597 P.2d 977 (Alaska 1979); City of Chicago v. Mayer, 56
Ill.2d 366, 308 N.E.2d 601 (1974); State v. Kee, 398 A.2d 384 (Me.1979); State
v. Caswell, 771 A.2d 375 (Me.2001); State v. Jacobs, 371 So.2d 801 (La.1979);
Anthony M. Dillof, Unraveling Unknowing Justification, 77 notre dame l .
rev . 1547 (2002).
109. United States v. Contento-Pachon, 723 F.2d 691 (9th Cir.1984);
United States v. Bailey, 444 U.S. 394, 100 S.Ct. 624, 62 L.Ed.2d 575 (1980);
Hunt v. State, 753 So.2d 609 (Fla.App.2000); State v. Anthuber, 201 Wis.2d
512, 549 N.W.2d 477 (App.1996).
110. State v. Fee, 126 N.H. 78, 489 A.2d 606 (1985); United States v. Sued-
Jimenez, 275 F.3d 1 (1st Cir.2001); United States v. Dorrell, 758 F.2d 427 (9th
Cir.1985); State v. Marley, 54 Haw. 450, 509 P.2d 1095 (1973); State v. Dan
singer, 521 A.2d 685 (Me.1987); State v. Champa, 494 A.2d 102 (R.I.1985);
Wilson v. State, 777 S.W.2d 823 (Tex.App.1989); State v. Cram, 157 Vt. 466,
600 A.2d 733 (1991).
111. United States v. Maxwell, 254 F.3d 21 (1st Cir.2001); Andrews v.
People, 800 P.2d 607 (Colo.1990); State v. Howley, 128 Idaho 874, 920 P.2d
391 (1996); State v. Dansinger, 521 A.2d 685 (Me.1987); Commonwealth v.
Leno, 415 Mass. 835, 616 N.E.2d 453 (1993); Commonwealth v. Lindsey,
396 Mass. 840, 489 N.E.2d 666 (1986); People v. Craig, 78 N.Y.2d 616, 578
N.Y.S.2d 471, 585 N.E.2d 783 (1991); State v. Warshow, 138 Vt. 22, 410 A.2d
1000 (1979).
113. Id.
114. John Lawrence Hill, A Utilitarian Theory of Duress, 84 iowa l . rev .
275 (1999); Rollin M. Perkins, Impelled Perpetration Restated, 33 hastings
l . j . 403 (1981); United States v. Johnson, 956 F.2d 894 (9th Cir.1992); Sanders
v. State, 466 N.E.2d 424 (Ind.1984); State v. Daoud, 141 N.H. 142, 679 A.2d
577 (1996); Alford v. State, 866 S.W.2d 619 (Tex.Crim.App.1993).
115. McGrowther, (1746) 18 How. St. Tr. 394.
116. United States v. LaFleur, 971 F.2d 200 (9th Cir.1991); Hunt v. State,
753 So.2d 609 (Fla.App.2000); Taylor v. State, 158 Miss. 505, 130 So. 502
(1930); State v. Finnell, 101 N.M. 732, 688 P.2d 769 (1984); State v. Nar
gashian, 26 R.I. 299, 58 A. 953 (1904); State v. Rocheville, 310 S.C. 20, 425
S.E.2d 32 (1993); Arp v. State, 97 Ala. 5, 12 So. 301 (1893).
117. State v. Nargashian, 26 R.I. 299, 58 A. 953 (1904).
118. People v. Merhige, 212 Mich. 601, 180 N.W. 418 (1920); People v.
Pantano, 239 N.Y. 416, 146 N.E. 646 (1925); Tully v. State, 730 P.2d 1206
(Okl.Crim.App.1986); Pugliese v. Commonwealth, 16 Va.App. 82, 428 S.E.2d
16 (1993).
119. United States v. Bakhtiari, 913 F.2d 1053 (2nd Cir.1990); R.I. Recre
ation Center v. Aetna Cas. & Surety Co., 177 F.2d 603 (1st Cir.1949); Sam v.
Commonwealth, 13 Va.App. 312, 411 S.E.2d 832 (1991).
120. Commonwealth v. Perl, 50 Mass.App.Ct. 445, 737 N.E.2d 937 (2000);
United States v. Contento-Pachon, 723 F.2d 691 (9th Cir.1984); State v. Ellis,
232 Or. 70, 374 P.2d 461 (1962); State v. Torphy, 78 Mo.App. 206 (1899).
121. People v. Richards, 269 Cal.App.2d 768, 75 Cal.Rptr. 597 (1969);
United States v. Bailey, 444 U.S. 394, 100 S.Ct. 624, 62 L.Ed.2d 575 (1980);
United States v. Gomez, 81 F.3d 846 (9th Cir.1996); United States v. Arthurs,
73 F.3d 444 (1st Cir.1996); United States v. Lee, 694 F.2d 649 (11th Cir.1983);
United States v. Campbell, 675 F.2d 815 (6th Cir.1982); State v. Daoud, 141
N.H. 142, 679 A.2d 577 (1996).
122. United States v. Bailey, 444 U.S. 394, 100 S.Ct. 624, 62 L.Ed.2d 575
(1980); People v. Handy, 198 Colo. 556, 603 P.2d 941 (1979); State v. Reese,
272 N.W.2d 863 (Iowa 1978); State v. Reed, 205 Neb. 45, 286 N.W.2d 111
(1979).
123. Fitzpatrick, [1977] N.I. 20; Hasan, [2005] U.K.H.L. 22, [2005] 4 All
E.R. 685, [2005] 2 Cr. App. Rep. 314, [2006] Crim. L.R. 142, [2005] All E.R.
(D) 299.
125. Michael A. Musmanno, Are Subordinate Officials Penally Responsi-
ble for Obeying Superior Orders which Direct Commission of Crime?, 67 dick .
l . rev . 221 (1963).
126. Axtell, (1660) 84 E.R. 1060; Calley v. Callaway, 519 F.2d 184 (5th
Cir.1975); United States v. Calley, 48 C.M.R. 19, 22 U.S.C.M.A. 534 (1973).
127. a . p . rogers , law on the battlefield 143 – 147 (1996).
128. Jurco v. State, 825 P.2d 909 (Alaska App.1992); State v. Stoehr, 134
Wis.2d 66, 396 N.W.2d 177 (1986).
129. For the machine learning, see section 1.1.2.
130. See, for example, sections 5.3.1, 5.3.2, and 5.3.3.
131. Vashon R. Rogers, Jr., De Minimis Non Curat Lex, 21 albany l . j . 186
(1880); Max L. Veech & Charles R. Moon, De Minimis non Curat Lex, 45 mich .
l . rev . 537 (1947).
132. The American Law Institute, Model Penal Code — Official Draft and
Explanatory Notes 40 (1962, 1985).
133. Stanislaw Pomorski, On Multiculturalism, Concepts of Crime, and
the “De Minimis” Defense, 1997 b.y.u. l. rev. 51 (1997).
6. Sentencing ai
1. bronislaw malinowski , crime and custom in savage society (1959,
1982).
2. exodus 21:24.
3. gertrude ezorsky , philosophical perspectives on punishment 102 – 134
(1972).
4. Sheldon Glueck, Principles of a Rational Penal Code, 41 harv . l . rev .
453 (1928).
5. Jackson Toby, Is Punishment Necessary?, 55 j . crim . l . criminology &
police sci . 332 (1964); C. G. Schoenfeld, In Defence of Retribution in the Law,
35 psychoanalytic q . 108 (1966).
6. francis allen , the decline of the rehabilitative ideal 66 (1981).
7. barbara hudson , understanding justice : an introduction to ideas ,
perspectives and controversies in modern penal theory 39 (1996, 2003);
jessica mitford , kind and usual punishment : the prison business (1974).
8. Russell L. Christopher, Deterring Retributivism: The Injustice of “Just”
Punishment, 96 nw . u . l . rev . 843 (2002); Douglas Husak, Holistic Retribu-
tion, 88 cal . l . rev . 991 (2000); Douglas Husak, Retribution in Criminal
Theory, 37 san diego l . rev . 959 (2000); Dan Markel, Are Shaming Punish-
ments Beautifully Retributive? Retributivism and the Implications for the
Alternative Sanctions Debate, 54 vand . l . rev . 2157 (2001).
9. andrew von hirsch , doing justice : the choice of punishment 74 – 7 5
(1976); Andrew von Hirsch, Proportionate Sentences: A Desert Perspective,
in principled sentencing : readings on theory and policy 115 (Andrew von
Hirsch, Andrew Ashworth, & Julian Roberts eds., 3d ed. 2009).
10. Paul H. Robinson & John M. Darley, The Utility of Desert, 91 nw . u . l .
rev . 453 (1997); Samuel Scheffler, Justice and Desert in Liberal Theory, 88
cal . l . rev . 965 (2000); Edward M. Wise, The Concept of Desert, 33 wayne l .
rev . 1343 (1987).
11. cesare beccaria , traité des délits et des peines (1764).

12. leon radzinowicz , a history of english criminal law and its ad -
ministration from 1750 vol . 1: the movement for reform (1948); jeremy
bentham , an introduction to the principles of morals and legislation ch.
13 (1789, 1996); Jeremy Bentham, Punishment and Deterrence, in principled
sentencing : readings on theory and policy 53 – 56 (Andrew von Hirsch,
Andrew Ashworth, & Julian Roberts eds., 3d ed. 2009); paul johann anselm
feuerbach , lehrbuch des gemeinen in deutschland gültigen peinlichen
rechts 117 (1812, 2007).
13. Johannes Andenaes, The General Preventive Effects of Punishment,
114 u . pa . l . rev . 949 (1966); Johannes Andenaes, The Morality of Deter-
rence, 37 u . chi . l . rev . 649 (1970).
14. james q . wilson , thinking about crime 123 – 142 (2d ed. 1985); Lau
rence H. Ross, Deterrence Regained: The Cheshire Constabulary’s “Breath-
alyser Blitz,” 6 j . legal stud . 241 (1977).
15. Gabriel Hallevy, The Recidivist Wants to Be Punished — Punishment
as an Incentive to Re-offend, 5 int ’ l j . punishment & sentencing 124 (2009).
16. leon radzinowicz & roger hood , a history of english criminal law
and its administration from 1750 vol . 5: the emergence of penal policy
8 – 57 (1986).
17. Andrew Ashworth, Rehabilitation, in principled sentencing : readings
on theory and policy 1, 1 – 10 (Andrew von Hirsch, Andrew Ashworth, &
Julian Roberts eds., 3d ed. 2009).
18. Robert Martinson, What Works? Questions and Answers about Prison
Reform, 35 public interest 22 (1974).
19. douglas s . lipton , robert martinson , & judith wilks , the effective -
ness of correctional treatment : a survey of treatment evaluation stud -
ies (1975).
20. lee sechrest , susan o . white , & elizabeth d . brown , the rehabilita -
tion of criminal offenders : problems and prospects 27 – 34 (1979); David F.
Greenberg, The Corrective Effects of Corrections: A Survey of Evaluation, in
corrections and punishment 111 (David F. Greenberg ed., 1977).
21. See, for example, michel foucault , discipline and punish : the birth
of the prison (1977); anthony m . platt , the child savers : the invention of
delinquency (2d ed. 1969, 1977); francis allen , the decline of the rehabili -
tative ideal 66 (1981); a clockwork orange (Warner Bros. 1971).
22. Gabriel Hallevy, Therapeutic Victim-Offender Mediation within the
Criminal Justice Process — Sharpening the Evaluation of Personal Potential
for Rehabilitation while Righting Wrongs under the Alternative-Dispute-
Resolution (ADR) Philosophy, 16 harv . negot . l . rev . 65 (2011).
23. david p . farrington & brandon c . welsh , preventing crime : what
works for children , offenders , victims and places (2006); lawrence w .
sherman , david p . farrington , doris leyton mackenzie , & brandon c .
welsh , evidence -b ased crime prevention (2006); rosemary sheehan , gill
mclvor , & chris trotter , what works with women offenders (2007);
Laaman v. Helgemoe, 437 F.Supp. 269 (1977).
24. Richard P. Seiter & Karen R. Kadela, Prisoner Reentry: What
Works,What Does Not, and What Is Promissing, 49 crime & delinquency 360
(2003); Clive R. Hollin, Treatment Programs for Offenders, 22 int ’ l j . of law
& psychiatry 361 (1999).
26. david garland , the culture of control : crime and social order in
contemporary society 102 (2002).
27. Ledger Wood, Responsibility and Punishment, 28 am . inst . crim . l . &
criminology 630, 639 (1938).
28. Martin P. Kafka, Sex Offending and Sexual Appetite: The Clinical and
Theoretical Relevance of Hypersexual Desire, 47 int ’ l j . of offender ther -
apy & comp . criminology 439 (2003); Matthew Jones, Overcoming the Myth
of Free Will in Criminal Law: The True Impact of the Genetic Revolution, 52
duke l . j . 1031 (2003); Sanford H. Kadish, Excusing Crime, 75 cal . l . rev . 257
(1987).
30. Stuart Field & Nico Jorg, Corporate Liability and Manslaughter:
Should We Be Going Dutch?, [1991] crim . l . r . 156 (1991).
31. Gerard E. Lynch, The Role of Criminal Law in Policing Corporate Mis-
conduct, 60 law & contemp . probs . 23 (1997); Richard Gruner, To Let the
Punishment Fit the Organization: Sanctioning Corporate Offenders Through
Corporate Probation, 16 am . j . crim . l . 1 (1988); Steven Walt & William S.
Laufer, Why Personhood Doesn’t Matter: Corporate Criminal Liability and
Sanctions, 18 am . j . crim . l . 263 (1991).
32. United States v. Allegheny Bottling Company, 695 F.Supp. 856
(1988).
33. Id. at p. 858.
34. Id.
35. John C. Coffee, Jr., “No Soul to Damn: No Body to Kick”: An Unscan-
dalised Inquiry Into the Problem of Corporate Punishment, 79 mich . l . rev .
386 (1981); steven box , power , crime and mystification 16 – 79 (1983); Brent
Fisse & John Braithwaite, The Allocation of Responsibility for Corporate
Crime: Individualism, Collectivism and Accountability, 11 sydney l . rev .
468 (1988).
36. Allegheny Bottling Company case, supra note 32, at p. 861.
37. Id. at p. 861.
38. russ versteeg , early mesopotamian law 126 (2000); g . r . driver &
john c . miles , the babylonian laws , vol . i : legal commentary 206, 495 – 496
(1952): “The capital penalty is most often expressed by saying that the of
fender ‘shall be killed’ . . . ; this occurs seventeen times in the first thirty-four
sections. A second form of expression, which occurs five times, is that ‘they
shall kill’ . . . the offender.”
39. Frank E. Hartung, Trends in the Use of Capital Punishment, 284(1)
annals am . acad . pol . & soc . sci . 8 (1952).
40. Gregg v. Georgia, 428 U.S. 153, S.Ct. 2909, 49 L.Ed.2d 859 (1979).
41. Provenzano v. Moore, 744 So.2d 413 (Fla. 1999); Dutton v. State, 123
Md. 373, 91 A. 417 (1914); Campbell v. Wood, 18 F.3d 662 (9th Cir. 1994);
Wilkerson v. Utah, 99 U.S. (9 Otto) 130, 25 L.Ed. 345 (1878); People v. Daugh
erty, 40 Cal.2d 876, 256 P.2d 911 (1953); Gray v. Lucas, 710 F.2d 1048 (5th
Cir. 1983); Hunt v. Nuth, 57 F.3d 1327 (4th Cir. 1995).
42. robert m . bohm , deathquest : an introduction to the theory and
practice of capital punishment in the united states 74 (1999).
43. Peter Fitzpatrick, “Always More to Do”: Capital Punishment and the
(De)Composition of Law, in the killing state — capital punishment in law ,
politics , and culture 117 (Austin Sarat ed., 1999); Franklin E. Zimring, The
Executioner’s Dissonant Song: On Capital Punishment and American Legal
Values, in the killing state — capital punishment in law , politics , and cul -
ture 137 (Austin Sarat ed., 1999).
44. Anne Norton, After the Terror: Mortality, Equality, Fraternity, in
the killing state — capital punishment in law , politics , and culture 27
(Austin Sarat ed., 1999); Hugo Adam Bedau, Abolishing the Death Penalty
Even for the Worst Murderers, in the killing state — capital punishment in
law , politics , and culture 40 (Austin Sarat ed., 1999).
45. Sean McConville, The Victorian Prison: England 1865 – 1965, in the
oxford history of the prison 131 (Norval Morris & David J. Rothman eds.,
1995); thorsten j . sellin , slavery and the penal system (1976); horsfall j .
turner , the annals of the wakefield house of corrections for three hun -
dred years 154 – 172 (1904).
46. john howard , the state of prisons in england and wales (1777,
1996).
47. David J. Rothman, For the Good of All: The Progressive Tradition in
Prison Reform, in history and crime : implications for criminal justice
policy 271 (James A. Inciardi & Charles E. Faupel eds., 1980).
48. Roy D. King, The Rise and Rise of Supermax: An American Solution
in Search of a Problem?, 1 punishment & society 163 (1999); chase riveland ,
supermax prisons : overview and general considerations (1999); jamie fell -
ner & joanne mariner , cold storage : super -m aximum security confinement
in indiana (1997).
49. dorris layton mackanzie & eugene e . hebert , correctional boot
camps : a tough intermediate sanction (1996); Sue Frank, Oklahoma Camp
Stresses Structure and Discipline, 53 corrections today 102 (1991); roberta
c . cronin , boot camps for adult and juvenile offenders : overview and
update (1994).

51. marc ancel , suspended sentence 14 – 17 (1971); Marc Ancel, The
System of Conditional Sentence or Sursis, 80 l . q . rev . 334, 336 (1964).
52. United Nations, probation and related measures , UN Department
of Social Affairs 29 – 30 (1951).
53. david j . rothman , conscience and convenience : the asylum and its
alternatives in progressive america (1980); frank schmalleger , criminal
justice today : an introductory text for the 21 st century 454 (2003).
54. Paul W. Keve, The Professional Character of the Presentence Report,
26 fed . probation 51 (1962).
55. harry e . allen , eric w . carlson , & evalyn c . parks , critical issues in
adult probation (1979); Crystal A. Garcia, Using Palmer’s Global Approach
to Evaluate Intensive Supervision Programs: Implications for Practice, 4 cor -
rect . manag . q . 60 (2000); andrew wright , gwyneth boswell , & martin
davies , contemporary probation practice (1993); michael cavadino & james
dignan , the penal system : an introduction 137 – 140 (2002).
56. John Harding, The Development of the Community Service, in alter -
native strategies for coping with crime 164 (Norman Tutt ed., 1978); Home
Office, review of criminal justice policy (1977); Ashlee Willis, Community
Service as an Alternative to Imprisonment: A Cautionary View, 24 probation
j . 120 (1977).
57. Julie Leibrich, Burt Galaway, & Yvonne Underhill, Community Sen-
tencing in New Zealand: A Survey of Users, 50 fed . probation 55 (1986).
58. James Austin & Barry Krisberg, The Unmet Promise of Alternatives,
28 j . res . crime & delinq . 374 (1982); Mark S. Umbreit, Community Service
Sentencing: Jail Alternatives or Added Sanction?, 45 fed . probation 3
(1981).
59. fiori rinaldi , imprisonment for non -payment of fines (1976); ger -
hardt grebing , the fine in comparative law : a survey of 21 countries
(1982).
60. leon radzinowicz & roger hood , a history of english criminal law
and its administration from 1750 vol . 5: the emergence of penal policy
(1986); peter young , punishment , money and the legal order : an analysis
of the emergence of monetary sanctions with special reference to scot -
land (1987).
61. Gail S. Funke, The Economics of Prison Crowding, 478 annals am .
acad . pol . & soc . sci . 86 (1985); Thomas Mathiesen, The Viewer Society:
Michel Foucault’s ‘Panopticon’ Revisited, 1 theoretical criminology 215
(1997).
62. sally t . hillsman & silvia s . g . casale , enforcement of fines as
criminal sanctions : the english experience and its relevance to american
practice (1986); Judith A. Greene, Structuring Criminal Fines: Making an
‘Intermediate Penalty’ More Useful and Equitable, 13 justice system j . 37
(1988); nigel walker & nicola padfield , sentencing : theory , law and prac -
tice (1996).
63. michael h . tonry & kathleen hatlestad , sentencing reform in over -
crowded times : a comparative perspective (1997).
65. United States v. Allegheny Bottling Company, 695 F.Supp. 856 (1988).
Selected Bi bliography
franz alexander & hugo staub , the criminal , the judge , and the public
(1931)
franz alexander , our age of unreason : a study of the irrational forces
in social life (rev. ed. 1971)
harry e . allen , eric w . carlson , & evalyn c . parks , critical issues in adult
probation (1979)
Susan M. Allan, No Code Orders v. Resuscitation: The Decision to Withhold
Life-Prolonging Treatment from the Terminally Ill, 26 wayne l . rev . 139
(1980)
Peter Alldridge, The Doctrine of Innocent Agency, 2 crim . l . f . 45 (1990)
francis allen , the decline of the rehabilitative ideal (1981)
Marc Ancel, The System of Conditional Sentence or Sursis, 80 l . q . rev . 334
(1964)
marc ancel , suspended sentence (1971)
Johannes Andenaes, The General Preventive Effects of Punishment, 114
u . pa . l . rev . 949 (1966)
Johannes Andenaes, The Morality of Deterrence, 37 u . chi . l . rev . 649
(1970)
Susan Leigh Anderson, Asimov’s “Three Laws of Robotics” and Machine Me-
taethics, 22 ai soc . 477 (2008)
Edward B. Arnolds & Norman F. Garland, The Defense of Necessity in Crimi-
nal Law: The Right to Choose the Lesser Evil, 65 j . crim . l . & criminology
289 (1974)
Andrew Ashworth, The Scope of Criminal Liability for Omissions, 84 l . q .
rev . 424 (1989)
Andrew Ashworth, Testing Fidelity to Legal Values: Official Involvement and
Criminal Justice , 63 mod . l . rev . 663 (2000)
andrew ashworth , principles of criminal law (5th ed. 2006)
Andrew Ashworth, Rehabilitation, in principled sentencing : readings on
theory and policy 1 (Andrew von Hirsch, Andrew Ashworth, & Julian
Roberts eds., 3d ed. 2009)
isaac asimov , i , robot (1950)
issac asimov , the rest of robots (1964)
Tom Athanasiou, High-Tech Politics: The Case of Artificial Intelligence, 92
socialist rev . 7 (1987)
213
Selected Bi bliography . 214
James Austin & Barry Krisberg, The Unmet Promise of Alternatives, 28 j . res .
crime & delinq . (1982)
john austin , the province of jurisprudence determined (1832, 2000)
bernard baars , in the theatre of consciousness (1997)
John S. Baker, Jr., State Police Powers and the Federalization of Local Crime,
72 temp . l . rev . 673 (1999)
stephen baker , final jeopardy : man vs . machine and the quest to know
everything (2011)
cesare beccaria , traité des délits et des peines (1764)
Hugo Adam Bedau, Abolishing the Death Penalty Even for the Worst Murder-
ers, in the killing state — capital punishment in law , politics , and cul -
ture 40 (Austin Sarat ed., 1999)
richard e . bellman , an introduction to artificial intelligence : can com -
puters think ? (1978)
jeremy bentham , an introduction to the principles of morals and legisla -
tion (1789, 1996)
Jeremy Bentham, Punishment and Deterrence, in principled sentencing :
readings on theory and policy (Andrew von Hirsch, Andrew Ashworth,
& Julian Roberts eds., 3d ed. 2009)
john biggs , the guilty mind (1955)
Ned Block, What Intuitions About Homunculi Don’t Show, 3 behavioral &
brain sci . 425 (1980)
robert m . bohm , deathquest : an introduction to the theory and practice
of capital punishment in the united states (1999)
Addison M. Bowman, Narcotic Addiction and Criminal Responsibility under
Durham, 53 geo . l . j . 1017 (1965)
steven box , power , crime and mystification (1983)
richard b . brandt , ethical theory (1959)
Kathleen F. Brickey, Corporate Criminal Accountability: A Brief History and
an Observation, 60 wash . u . l . q . 393 (1983)
Bruce Bridgeman, Brains + Programs = Minds, 3 behavioral & brain sci . 427
(1980)
walter bromberg , from shaman to psychotherapist : a history of the
treatment of mental illness (1975)
Andrew G. Brooks & Ronald C. Arkin, Behavioral Overlays for Non-Verbal
Communication Expression on a Humanoid Robot, 22 auton . robots 55
(2007)
Timothy L. Butler, Can a Computer Be an Author — Copyright Aspects of Ar-
tificial Intelligence, 4 comm . ent . l . s . 707 (1982)
Kenneth L. Campbell, Psychological Blow Automatism: A Narrow Defence,
23 crim . l . q . 342 (1981)
W. G. Carson, Some Sociological Aspects of Strict Liability and the Enforce-
ment of Factory Legislation, 33 mod . l . rev . 396 (1970)
W. G. Carson, The Conventionalisation of Early Factory Crime, 7 int ’ l j . of
sociology of law 37 (1979)
Derrick Augustus Carter, Bifurcations of Consciousness: The Elimination of

the Self-Induced Intoxication Excuse, 64 mo . l . rev . 383 (1999)
michael cavadino & james dignan , the penal system : an introduction
(2002)
eugene charniak & drew mcdermott , introduction to artificial intelli -
gence (1985)
Russell L. Christopher, Deterring Retributivism: The Injustice of “Just” Pun-
ishment, 96 nw . u . l . rev . 843 (2002)
arthur c . clarke , 2001: a space odyssey (1968)
John C. Coffee, Jr., “No Soul to Damn: No Body to Kick”: An Unscandalised
Inquiry into the Problem of Corporate Punishment, 79 mich . l . rev . 386
(1981)
sir edward coke , institutions of the laws of england — third part (6th ed.
1681, 1817, 2001)
Dana K. Cole, Expending Felony-Murder in Ohio: Felony-Murder or Murder-
Felony, 63 ohio st . l . j . 15 (2002)
roberta c . cronin , boot camps for adult and juvenile offenders : overview
and update (1994)
George R. Cross & Cary G. Debessonet, An Artificial Intelligence Application
in the Law: CCLIPS, A Computer Program that Processes Legal Informa-
tion, 1 high tech . l . j . 329 (1986)
Homer D. Crotty, The History of Insanity as a Defence to Crime in English
Common Law, 12 cal . l . rev . 105 (1924)
michael dalton , the countrey justice (1618, 2003)
Donald Davidson, Turing’s Test, in modelling the mind (K. A. Mohyeldin
Said, W. H. Newton-Smith, R. Viale, & K. V. Wilkes eds., 1990)
Michael J. Davidson, Feminine Hormonal Defenses: Premenstrual Syndrome
and Postpartum Psychosis, 5 army lawyer (2000)
Richard Delgado, Ascription of Criminal States of Mind: Toward a Defense
Theory for the Coercively Persuaded (“Brainwashed”) Defendant, 63 minn .
l . rev . 1 (1978)
daniel c . dennett , brainstorms (1978)
daniel c . dennett , the intentional stance (1987)
Daniel C. Dennett, Evolution, Error, and Intentionality, in the foundations
of artificial intelligence 190 (Derek Pertridge & Yorick Wilks eds., 1990,
2006)
rené descartes , discours de la méthode pour bien conduire sa raison et
chercher la vérité dans les sciences (1637)
Anthony M. Dillof, Unraveling Unknowing Justification, 77 notre dame l .
rev . 1547 (2002)
Dolores A. Donovan & Stephanie M. Wildman, Is the Reasonable Man Obso-
lete? A Critical Perspective on Self-Defense and Provocation, 14 loy . l . a .
l . rev . 435, 441 (1981)
aage gerhardt drachmann , the mechanical technology of greek and
roman antiquity : a study of the literary sources (1963)
Joshua Dressler, Professor Delgado’s “Brainwashing” Defense: Courting a De-

terminist Legal System, 63 minn . l . rev . 335 (1978)
Joshua Dressler, Rethinking Heat of Passion: A Defense in Search of a Ratio-
nale, 73 j . crim . l . & criminology 421 (1982)
Joshua Dressler, Battered Women Who Kill Their Sleeping Tormenters: Re-
flections on Maintaining Respect for Human Life while Killing Moral Mon-
sters, in criminal law theory — doctrines of the general part 259 (Ste
phen Shute & A. P. Simester eds., 2005)
g . r . driver & john c . miles , the babylonian laws , vol . i : legal commen -
tary (1952)
antony robin duff , criminal attempts (1996)
Fernand N. Dutile & Harold F. Moore, Mistake and Impossibility: Arranging
Marriage Between Two Difficult Partners, 74 nw . u . l . rev . 166 (1980)
Justice Ellis, Criminal Law as an Instrument of Social Control, 17 victoria u .
wellington l . rev . 319 (1987)
gertrude ezorsky , philosophical perspectives on punishment (1972)
Judith Fabricant, Homicide in Response to a Threat of Rape: A Theoretical
Examination of the Rule of Justification, 11 golden gate u . l . rev . 945
(1981)
david p . farrington & brandon c . welsh , preventing crime : what works for
children , offenders , victims and places (2006)
edward a . feigenbaum & pamela mccorduck , the fifth generation : artifi -
cial intelligence and japan ’ s computer challenge to the world (1983)
S. Z. Feller, Les Délits de Mise en Danger, 40 rev . int . de droit pénal 179
(1969)
jamie fellner & joanne mariner , cold storage : super -m aximum security
confinement in indiana (1997)
Robert P. Fine & Gary M. Cohen, Is Criminal Negligence a Defensible Basis
for Criminal Liability?, 16 buff . l . rev . 749 (1966)
paul johann anselm feuerbach , lehrbuch des gemeinen in deutschland
gültigen peinlichen rechts (1812, 2007)
Stuart Field & Nico Jorg, Corporate Liability and Manslaughter: Should We
Be Going Dutch?, [1991] crim . l . r . 156 (1991)
Herbert Fingarette, Addiction and Criminal Responsibility, 84 yale l . j . 413
(1975)
arthur e . fink , causes of crime : biological theories in the united states ,
1800 – 1 915 (1938)
john finnis , natural law and natural rights (1980)
Brent Fisse & John Braithwaite, The Allocation of Responsibility for Corpo-
rate Crime: Individualism, Collectivism and Accountability, 11 sydney l .
rev . 468 (1988)
Peter Fitzpatrick, “Always More to Do”: Capital Punishment and the (De)
Composition of Law, in the killing state — capital punishment in law ,
politics , and culture 117 (Austin Sarat ed., 1999)
owen j . flanagan , jr ., the science of the mind (2d ed. 1991)

george p . fletcher , rethinking criminal law (1978, 2000)
George P. Fletcher, The Nature of Justification, in action and value in crimi -
nal law 175 (Stephen Shute, John Gardner, & Jeremy Horder eds., 2003)
Jerry A. Fodor, Modules, Frames, Fridgeons, Sleeping Dogs and the Music
of the Spheres, in the robot ’ s dilemma : the frame problem in artificial
intelligence (Zenon W. Pylyshyn ed., 1987)
Keith Foren, Casenote: In Re Tyvonne M. Revisited: The Criminal Infancy
Defense in Connecticut, 18 q . l . rev . 733 (1999)
michel foucault , madness and civilization (1965)
michel foucault , discipline and punish : the birth of the prison (1977)
Sue Frank, Oklahoma Camp Stresses Structure and Discipline, 53 correc -
tions today 102 (1991)
Lionel H. Frankel, Criminal Omissions: A Legal Microcosm, 11 wayne l . rev .
367 (1965)
Lionel H. Frankel, Narcotic Addiction, Criminal Responsibility and Civil
Commitment, 1966 utah l . rev . 581 (1966)
Robert M. French, Subcognition and the Limits of the Turing Test, 99 mind 53
(1990)
K. W. M. Fulford, Value, Action, Mental Illness, and the Law, in action
and value in criminal law 279 (Stephen Shute, John Gardner, & Jeremy
Horder eds., 2003)
Gail S. Funke, The Economics of Prison Crowding, 478 annals am . acad .
pol . & soc . sci . 86 (1985)
Jonathan M. E. Gabbai, Complexity and the Aerospace Industry: Understand
ing Emergence by Relating Structure to Performance Using Multi-Agent
Systems (PhD Thesis, University of Manchester, 2005)
Crystal A. Garcia, Using Palmer’s Global Approach to Evaluate Intensive Su-
pervision Programs: Implications for Practice, 4 correction manag . q . 60
(2000)
howard gardner , the mind ’ s new science : a history of the cognitive revo -
lution (1985)
david garland , the culture of control : crime and social order in contem -
porary society (2002)
Chas E. George, Limitation of Police Powers, 12 law . & banker & s . bench &
b . rev . 740 (1919)
Jack P. Gibbs, A Very Short Step toward a General Theory of Social Control,
1985 am . b . found res . j . 607 (1985)
sander l . gilman , seeing the insane (1982)
P. R. Glazebrook, Criminal Omissions: The Duty Requirement in Offences
Against the Person, 55 l . q . rev . 386 (1960)
robert m . glorioso & fernando c . colon osorio , engineering intelligent
systems : concepts and applications (1980)
Sheldon Glueck, Principles of a Rational Penal Code, 41 harv. l. rev. 453 (1928)
sir gerald gordon , the criminal law of scotland (1st ed. 1967)
gerhardt grebing , the fine in comparative law : a survey of 21 countries
(1982)
Kent Greenawalt, The Perplexing Borders of Justification and Excuse, 84
colum . l . rev . 1897 (1984)
Kent Greenawalt, Distinguishing Justifications from Excuses, 49 law & con -
temp . probs . 89 (1986)
David F. Greenberg, The Corrective Effects of Corrections: A Survey of Evalu-
ation, in corrections and punishment 111 (David F. Greenberg ed., 1977)
Judith A. Greene, Structuring Criminal Fines: Making an ‘Intermediate Pen-
alty’ More Useful and Equitable, 13 justice system j . 37 (1988)
Richard Gruner, To Let the Punishment Fit the Organization: Sanctioning
Corporate Offenders Through Corporate Probation, 16 am . j . crim . l . 1
(1988)
jerome hall , general principles of criminal law (2d ed. 1960, 2005)
Jerome Hall, Intoxication and Criminal Responsibility, 57 harv . l . rev . 1045
(1944)
Jerome Hall, Negligent Behaviour Should Be Excluded from Penal Liability,
63 colum . l . rev . 632 (1963)
Seymour L. Halleck, The Historical and Ethical Antecedents of Psychiatric
Criminology, in psychiatric aspects of criminology 8 (Halleck & Brom
berg eds., 1968)
Gabriel Hallevy, The Recidivist Wants to Be Punished — Punishment as an
Incentive to Re-offend, 5 int ’ l j . punishment & sentencing 124 (2009)
gabriel hallevy , a modern treatise on the principle of legality in crimi -
nal law (2010)
Gabriel Hallevy, The Criminal Liability of Artificial Intelligence Entities —
from Science Fiction to Legal Social Control, 4 akron intell . prop . j . 171
(2010)
Gabriel Hallevy, Unmanned Vehicles — Subordination to Criminal Law under
the Modern Concept of Criminal Liability, 21 j . l . inf . & sci . 200 (2011)
Gabriel Hallevy, Therapeutic Victim-Offender Mediation within the Crimi-
nal Justice Process — Sharpening the Evaluation of Personal Potential
for Rehabilitation while Righting Wrongs under the Alternative-Dispute-
Resolution (ADR) Philosophy, 16 harv . negot . l . rev . 65 (2011)
gabriel hallevy , the matrix of derivative criminal liability (2012)
John Harding, The Development of the Community Service, in alternative
strategies for coping with crime 164 (Norman Tutt ed., 1978)
herbert l . a . hart , punishment and responsibility : essays in the philoso -
phy of law (1968)
Frank E. Hartung, Trends in the Use of Capital Punishment, 284(1) annals
am . acad . pol . & soc . sci . 8 (1952)
John Haugeland, Semantic Engines: An Introduction to Mind Design, in mind
design 1 (John Haugeland ed., 1981)
john haugeland , artificial intelligence : the very idea (1985)

pamela rae heath , the pk zone : a cross -c ultural review of psychokinesis
( pk ) (2003)
hermann von helmholtz , the facts of perception (1878)
John Lawrence Hill, A Utilitarian Theory of Duress, 84 iowa l . rev . 275 (1999)
sally t . hillsman & silvia s . g . casale , enforcement of fines as criminal
sanctions : the english experience and its relevance to american prac -
tice (1986)
Harold L. Hirsh & Richard E. Donovan, The Right to Die: Medico-Legal Impli-
cations of In re Quinlan, 30 rutgers l . rev . 267 (1977)
andrew von hirsch , doing justice : the choice of punishment (1976)
Andrew von Hirsch, Proportionate Sentences: A Desert Perspective, in
principled sentencing : readings on theory and policy 115 (Andrew von
Hirsch, Andrew Ashworth, & Julian Roberts eds., 3d ed. 2009)
W. H. Hitchler, Necessity as a Defence in Criminal Cases, 33 dick . l . rev . 138
(1929)
thomas hobbes , leviathan or the matter , forme and power of a common
wealth ecclesiasticall and civil (1651)
douglas r . hofstadter , gödel , escher , bach : an eternal golden braid
(1979, 1999)
William Searle Holdsworth, English Corporation Law in the 16th and 17th
Centuries, 31 yale l . j . 382 (1922)
william searle holdsworth , a history of english law (1923)
Winifred H. Holland, Automatism and Criminal Responsibility, 25 crim .
l . q . 95 (1982)
Clive R. Hollin, Treatment Programs for Offenders, 22 int ’ l j . of law &
psychiatry 361 (1999)
oliver w . holmes , the common law (1881, 1923)
Oliver W. Holmes, Agency, 4 harv . l . rev . 345 (1891)
henry holt , telekinesis (2005)
Morton J. Horwitz, The Rise and Early Progressive Critique of Objective Cau-
sation, in the politics of law : a progressive critique 471 (David Kairys
ed., 3d ed. 1998)
john howard , the state of prisons in england and wales (1777, 1996)
feng -h siung hsu , behind deep blue : building the computer that defeated
the world chess champion (2002)
barbara hudson , understanding justice : an introduction to ideas , per -
spectives and controversies in modern penal theory (1996, 2003)
Graham Hughes, Criminal Omissions, 67 yale l . j . 590 (1958)
bishop carleton hunt , the development of the business corporation in
england 1800 – 1 867 (1963)
Douglas Husak, Holistic Retribution, 88 cal . l . rev . 991 (2000)
Douglas Husak, Retribution in Criminal Theory, 37 san diego l . rev . 959
(2000)
Douglas Husak & Andrew von Hirsch, Culpability and Mistake of Law, in
action and value in criminal law 157 (Stephen Shute, John Gardner, &
Jeremy Horder eds., 2003)
Peter Barton Hutt & Richard A. Merrill, Criminal Responsibility and the Right
to Treatment for Intoxication and Alcoholism, 57 geo . l . j . 835 (1969)
ray jackendoff , consciousness and the computational mind (1987)
william james , the principles of psychology (1890)
phillip n . johnson -l aird , mental models (1983)
Matthew Jones, Overcoming the Myth of Free Will in Criminal Law: The True
Impact of the Genetic Revolution, 52 duke l . j . 1031 (2003)
Sanford Kadish, Respect for Life and Regard for Rights in the Criminal Law,
64 cal . l . rev . 871 (1976)
Sanford H. Kadish, Excusing Crime, 75 cal . l . rev . 257 (1987)
Martin P. Kafka, Sex Offending and Sexual Appetite: The Clinical and Theo-
retical Relevance of Hypersexual Desire, 47 int ’ l j . of offender therapy
& comp . criminology 439 (2003)
immanuel kant , our duties to animals (1780)
A. W. G. Kean, The History of the Criminal Liability of Children, 53 l . q . rev .
364 (1937)
vojislav kecman , learning and soft computing , support vector machines ,
neural networks and fuzzy logic models (2001)
Edwin R. Keedy, Ignorance and Mistake in the Criminal Law, 22 harv . l .
rev . 75 (1909)
Paul W. Keve, The Professional Character of the Presentence Report, 26 fed .
probation 51 (1962)
antony kenny , will , freedom and power (1975)
antony kenny , what is faith ? (1992)
Roy D. King, The Rise and Rise of Supermax: An American Solution in
Search of a Problem?, 1 punishment & society 163 (1999)
raymond kurzweil , the age of intelligent machines (1990)
nicola lacey & celia wells , reconstructing criminal law — critical per -
spectives on crime and the criminal process (2d ed. 1998)
nicola lacey , celia wells , & oliver quick , reconstructing criminal law
(3d ed. 2003, 2006)
j . g . landels , engineering in the ancient world (rev. ed. 2000)
William S. Laufer, Corporate Bodies and Guilty Minds, 43 emory l . j . 647
(1994)
gottfried wilhelm leibniz , characteristica universalis (1676)
Julie Leibrich, Burt Galaway, & Yvonne Underhill, Community Sentencing in
New Zealand: A Survey of Users, 50 fed . probation 55 (1986)
lawrence lessig , code and other laws of cyberspace (1999)
david levy , robots unlimited : life in a virtual age (2006)
david levy , love and sex with robots : the evolution of human -r obot rela -
tionships (2007)
david levy & monty newborn , how computers play chess (1991)
K. W. Lidstone, Social Control and the Criminal Law, 27 brit . j . criminology
31 (1987)
douglas s . lipton , robert martinson , & judith wilks , the effectiveness
of correctional treatment : a survey of treatment evaluation studies
(1975)
Frederick J. Ludwig, Rationale of Responsibility for Young Offenders, 29 neb .
l . rev . 521 (1950)
george f . luger , artificial intelligence : structures and strategies for
complex problem solving (2001)
george f . luger & william a . stubblefield , artificial intelligence : struc -
tures and strategies for complex problem solving (6th ed. 2008)
William G. Lycan, Introduction, in mind and cognition 3 (William G. Lycan
ed., 1990)
Gerard E. Lynch, The Role of Criminal Law in Policing Corporate Miscon-
duct, 60 law & contemp . probs . 23 (1997)
Peter Lynch, The Origins of Computer Weather Prediction and Climate Mod-
eling, 227 journal of computational physics 3431 (2008)
david lyons , forms and limits of utilitarianism (1965)
David Lyons, Open Texture and the Possibility of Legal Interpretation, 18 law
phil . 297 (1999)
dorris layton mackanzie & eugene e . hebert , correctional boot camps :
a tough intermediate sanction (1996)
bronislaw malinowski , crime and custom in savage society (1959, 1982)
david manners & tsugio makimoto , living with the chip (1995)
Dan Markel, Are Shaming Punishments Beautifully Retributive? Retributiv-
ism and the Implications for the Alternative Sanctions Debate, 54 vand .
l . rev . 2157 (2001)
Robert Martinson, What Works? Questions and Answers about Prison
Reform, 35 public interest 22 (1974)
Thomas Mathiesen, The Viewer Society: Michel Foucault’s ‘Panopticon’ Re-
visited, 1 theoretical criminology 215 (1997)
Peter McCandless, Liberty and Lunacy: The Victorians and Wrongful Con-
finement, in madhouses , mad -d octors , and madmen : the social history
of psychiatry in the victorian era (Scull ed., 1981)
Aileen McColgan, In Defence of Battered Women who Kill, 13 oxford j .
legal stud . 508 (1993)
Sean McConville, The Victorian Prison: England 1865 – 1965, in the oxford
history of the prison 131 (Norval Morris & David J. Rothman eds., 1995)
j . r . mcdonald , g . m . burt , j . s . zielinski , & s . d . j . mcarthur , intelligent
knowledge based system in electrical power engineering (1997)
colin mcginn , the problem of consciousness : essays towards a resolution
(1991)
karl menninger , martin mayman , & paul pruyser , the vital balance (1963)
Alan C. Michaels, Imposing Constitutional Limits on Strict Liability: Lessons

from the American Experience, in appraising strict liability 218 (A. P.
Simester ed., 2005)
donald michie & rory johnston , the creative computer (1984)
Justine Miller, Criminal Law — An Agency for Social Control, 43 yale l . j . 691
(1934)
marvin minsky , the society of mind (1986)
jessica mitford , kind and usual punishment : the prison business (1974)
Patrick Montague, Self-Defense and Choosing Between Lives, 40 phil . stud .
207 (1981)
michael moore , law and psychiatry : rethinking the relationship (1984)
George Mora, Historical and Theoretical Trends in Psychiatry, in comprehen -
sive textbook of psychiatry 1 (Alfred M. Freedman, Harold Kaplan, &
Benjamin J. Sadock eds., 2d ed. 1975)
hans moravec , robot : mere machine to transcendent mind (1999)
Norval Morris, Somnambulistic Homicide: Ghosts, Spiders, and North Kore-
ans, 5 res judicatae 29 (1951)
tim morris , computer vision and image processing (2004)
Gerhard O. W. Mueller, Mens Rea and the Corporation — A Study of the
Model Penal Code Position on Corporate Criminal Liability, 19 u . pitt . l .
rev . 21 (1957)
Michael A. Musmanno, Are Subordinate Officials Penally Responsible for
Obeying Superior Orders which Direct Commission of Crime?, 67 dick . l .
rev . 221 (1963)
monty newborn , deep blue (2002)
allen newell & herbert a . simon , human problem solving (1972)
edward norbeck , religion in primitive society (1961)
Anne Norton, After the Terror: Mortality, Equality, Fraternity, in the kill -
ing state — capital punishment in law , politics , and culture 27 (Austin
Sarat ed., 1999)
Scott T. Noth, A Penny for Your Thoughts: Post-Mitchell Hate Crime Laws
Confirm a Mutating Effect upon Our First Amendment and the Govern-
ment’s Role in Our Lives, 10 regent u . l . rev . 167 (1998)
david ormerod , smith & hogan criminal law (11th ed. 2005)
n . p . padhy , artificial intelligence and intelligent systems (2005, 2009)
william paley , a treatise on the law of principal and agent (2d ed. 1847)
dan w . patterson , introduction to artificial intelligence and expert sys -
tems (1990)
Monrad G. Paulsen, Intoxication as a Defense to Crime, 1961 u . ill . l . f . 1
(1961)
Rollin M. Perkins, Negative Acts in Criminal Law, 22 iowa l . rev . 659 (1937)
Rollin M. Perkins, Ignorance and Mistake in Criminal Law, 88 u . pa . l . rev .
35 (1940)
Rollin M. Perkins, “Knowledge” as a Mens Rea Requirement, 29 hastings l . j .
953 (1978)
Rollin M. Perkins, Impelled Perpetration Restated, 33 hastings l . j . 403

(1981)
anthony m . platt , the child savers : the invention of delinquency (2d ed.
1969, 1977)
Anthony Platt & Bernard L. Diamond, The Origins of the “Right and Wrong”
Test of Criminal Responsibility and Its Subsequent Development in the
United States: An Historical Survey, 54 cal . l . rev . 1227 (1966)
frederick pollock & frederick william maitland , the history of english
law before the time of edward i (rev. 2d ed. 1898)
Stanislaw Pomorski, On Multiculturalism, Concepts of Crime, and the “De
Minimis” Defense, 1997 b . y . u . l . rev . 51 (1997)
james cowles prichard , a treatise on insanity and other disorders affect -
ing the mind (1835)
gustav radbruch , der handlungsbegriff in seiner bedeutung für das straf -
rechtssystem (1904)
leon radzinowicz , a history of english criminal law and its administra -
tion from 1750 vol . 1: the movement for reform (1948)
leon radzinowicz & roger hood , a history of english criminal law and its
administration from 1750 vol . 5: the emergence of penal policy (1986)
Craig W. Reynolds, Herds and Schools: A Distributed Behavioral Model, 21
comput . graph . (1987)
elaine rich & kevin knight , artificial intelligence (2d ed. 1991)
fiori rinaldi , imprisonment for non -payment of fines (1976)
Edwina L. Rissland, Artificial Intelligence and Law: Stepping Stones to a
Model of Legal Reasoning, 99 yale l . j . 1957 (1990)
chase riveland , supermax prisons : overview and general considerations
(1999)
olivia f . robinson , the criminal law of ancient rome (1995)
Paul H. Robinson, A Theory of Justification: Societal Harm as a Prerequisite
for Criminal Liability, 23 u . c . l . a . l . rev . 266 (1975)
Paul H. Robinson & John M. Darley, The Utility of Desert, 91 nw . u . l . rev .
453 (1997)
Paul H. Robinson, Testing Competing Theories of Justification, 76 n . c . l . rev .
1095 (1998)
p . rogers , law on the battlefield (1996)
Vashon R. Rogers, Jr., De Minimis Non Curat Lex, 21 albany l . j . 186 (1880)
george rosen , madness in society : chapters in the historical sociology of
mental illness (1969)
Laurence H. Ross, Deterrence Regained: The Cheshire Constabulary’s
“Breathalyser Blitz,” 6 j . legal stud . 241 (1977)
david j . rothman , conscience and convenience : the asylum and its alter -
natives in progressive america (1980)
David J. Rothman, For the Good of All: The Progressive Tradition in Prison
Reform, in history and crime : implications for criminal justice policy
271 (James A. Inciardi & Charles E. Faupel eds., 1980)
claus roxin , strafrecht — allgemeiner teil i (4 Auf. 2006)

stuart j . russell & peter norvig , artificial intelligence : a modern ap -
proach (2002)
william oldnall russell , a treatise on crimes and misdemeanors (1843, 1964)
Cheyney C. Ryan, Self-Defense, Pacificism, and the Possibility of Killing, 93
ethics 508 (1983)
gilbert ryle , the concept of mind (1954)
Francis Bowes Sayre, Criminal Responsibility for the Acts of Another, 43
harv . l . rev . 689 (1930)
Francis Bowes Sayre, Mens Rea, 45 harv . l . rev . 974 (1932)
Francis Bowes Sayre, Public Welfare Offenses, 33 colum . l . rev . 55 (1933).
robert j . schalkoff , artificial intelligence : an engineering approach
(1990)
Roger C. Schank, What is ai , Anyway?, in the foundations of artificial in -
telligence 3 (Derek Pertridge & Yorick Wilks eds., 1990, 2006)
Samuel Scheffler, Justice and Desert in Liberal Theory, 88 cal . l . rev . 965
(2000)
G. Schoenfeld, In Defence of Retribution in the Law, 35 psychoanalytic q .
108 (1966)
frank schmalleger , criminal justice today : an introductory text for the
21 st century (2003)
william robert scott , the constitution and finance of english , scottish
and irish joint -s tock companies to 1720 (1912)
John R. Searle, Minds, Brains & Programs, 3 behavioral & brain sci . 417
(1980)
john r . searle , minds , brains and science (1984)
john r . searle , the rediscovery of mind (1992)
lee sechrest , susan o . white , & elizabeth d . brown , the rehabilitation of
criminal offenders : problems and prospects (1979)
Richard P. Seiter & Karen R. Kadela, Prisoner Reentry: What Works,What
Does Not, and What Is Promissing, 49 crime & delinquency 360 (2003)
thorsten j . sellin , slavery and the penal system (1976)
Robert N. Shapiro, Of Robots, Persons, and the Protection of Religious Be-
liefs, 56 s . cal . l . rev . 1277 (1983)
rosemary sheehan , gill mclvor , & chris trotter , what works with women
offenders (2007)
lawrence w . sherman , david p . farrington , doris leyton mackenzie , &
brandon c . welsh , evidence -b ased crime prevention (2006)
Nancy Sherman, The Place of the Emotions in Kantian Morality, in identity ,
character , and morality 145 (Owen Flanagan & Amelie O. Rotry eds.,
1990)
Stephen Shute, Knowledge and Belief in the Criminal Law, in criminal law
theory — doctrines of the general part 182 (Stephen Shute & A.P.
Simester eds., 2005)
R. U. Singh, History of the Defence of Drunkenness in English Criminal Law,

49 law q . rev . 528 (1933)
vieda skultans , english madness : ideas on insanity , 1580 – 1 890 (1979)
Aaron Sloman, Motives, Mechanisms, and Emotions, in the philosophy of
artificial intelligence 231 (Margaret A. Boden ed., 1990)
john j . c . smart & bernard williams , utilitarianism — for and against
(1973)
rudolph sohm , the institutes of roman law (3d ed. 1907)
Lawrence B. Solum, Legal Personhood for Artificial Intelligences, 70 n . c . l .
rev . 1231 (1992)
milan sonka , vaclav hlavac , & roger boyle , image processing , analysis ,
and machine vision (2008)
walter w . soroka , analog methods in computation and simulation (1954)
John R. Spencer & Antje Pedain, Approaches to Strict and Constructive Li-
ability in Continental Criminal Law, in appraising strict liability 237
(A. P. Simester ed., 2005)
Jane Stapelton, Law, Causation and Common Sense, 8 oxford j . legal stud .
111 (1988)
G. R. Sullivan, Knowledge, Belief, and Culpability, in criminal law
theory — doctrines of the general part 207 (Stephen Shute & A. P.
G. R. Sullivan, Strict Liability for Criminal Offences in England and Wales
Following Incorporation into English Law of the European Convention
on Human Rights, in appraising strict liability 195 (A. P. Simester ed.,
2005)
roger j . sullivan , immanuel kant ’ s moral theory (1989)
Victor Tadors, Recklessness and the Duty to Take Care, in criminal law
theory — doctrines of the general part 227 (Stephen Shute & A. P.
steven l . tanimoto , elements of artificial intelligence : an introduction
using lisp (1987)
Lawrence Taylor & Katharina Dalton, Premenstrual Syndrome: A New Crimi-
nal Defense?, 19 cal . w . l . rev . 269 (1983)
judith jarvis thomson , rights , restitution and risk : essays in moral
theory (1986)
benjamin thorpe , ancient laws and institutes of england (1840, 2004)
Lawrence P. Tiffany & Carl A. Anderson, Legislating the Necessity Defense in
Criminal Law, 52 denv . l . j . 839 (1975)
Janet A. Tighe, Francis Wharton and the Nineteenth Century Insanity De-
fense: The Origins of a Reform Tradition, 27 am . j . legal hist . 223 (1983)
Jackson Toby, Is Punishment Necessary?, 55 j . crim . l . criminology & police
sci . 332 (1964)
michael h . tonry & kathleen hatlestad , sentencing reform in over -
crowded times : a comparative perspective (1997)
Richard H. S. Tur, Subjectivism and Objectivism: Towards Synthesis, in

action and value in criminal law 213 (Stephen Shute, John Gardner, &
Jeremy Horder eds., 2003)
Alan Turing, Computing Machinery and Intelligence, 59 mind 433 (1950)
austin turk , criminality and legal order (1969)
horsfall j . turner , the annals of the wakefield house of corrections for
three hundred years (1904)
alan tyree , expert systems in law (1989)
Mark S. Umbreit, Community Service Sentencing: Jail Alternatives or Added
Sanction?, 45 fed . probation 3 (1981)
Max L. Veech & Charles R. Moon, De Minimis non Curat Lex, 45 mich . l . rev .
537 (1947)
russ versteeg , early mesopotamian law (2000)
John Barker Waite, The Law of Arrest, 24 tex . l . rev . 279 (1946)
nigel walker & nicola padfield , sentencing : theory , law and practice
(1996)
Andrew Walkover, The Infancy Defense in the New Juvenile Court, 31
u . c . l . a . l . rev . 503 (1984)
Steven Walt & William S. Laufer, Why Personhood Doesn’t Matter: Corporate
Criminal Liability and Sanctions, 18 am . j . crim . l . 263 (1991)
Mary Anne Warren, On the Moral and Legal Status of Abortion, in ethics in
practice (Hugh Lafollette ed., 1997)
donald a . waterman , a guide to expert systems (1986)
max weber , economy and society : an outline of interpretive sociology (1968)
henry weihofen , mental disorder as a criminal defense (1954)
Paul Weiss, On the Impossibility of Artificial Intelligence, 44 rev . metaphys -
ics 335 (1990)
Celia Wells, Battered Woman Syndrome and Defences to Homicide: Where
Now?, 14 legal stud . 266 (1994)
Yueh-Hsuan Weng, Chien-Hsun Chen, & Chuen-Tsai Sun, The Legal Crisis of
Next Generation Robots: On Safety Intelligence, Proceedings of the 11th
International Conference on Artificial Intelligence and Law 205 (2007)
Yueh-Hsuan Weng, Chien-Hsun Chen, & Chuen-Tsai Sun, Toward the
Human-Robot Co-Existence Society: On Safety Intelligence for Next Gen-
eration Robots, 1 int . j . soc . robot . 267 (2009)
francis antony whitlock , criminal responsibility and mental illness
(1963)
glanville williams , criminal law : the general part (2d ed. 1961)
Glanville Williams, Oblique Intention, 46 camb . l . j . 417 (1987)
Glanville Williams, The Draft Code and Reliance upon Official Statements, 9
legal stud . 177 (1989)
Glanville Williams, Innocent Agency and Causation, 3 crim . l . f . 289 (1992)
Ashlee Willis, Community Service as an Alternative to Imprisonment: A
Cautionary View, 24 probation j . 120 (1977)
edward o . wilson , sociobiology : the new synthesis (1975)

james q . wilson , thinking about crime (2d ed. 1985)
terry winograd & fernando c . flores , understanding computers and cog -
nition : a new foundation for design (1986, 1987)
Terry Winograd, Thinking Machines: Can There Be? Are We?, in the founda -
tions of artificial intelligence 167 (Derek Pertridge & Yorick Wilks eds.,
1990, 2006)
patrick henry winston , artificial intelligence (3d ed. 1992)
Edward M. Wise, The Concept of Desert, 33 wayne l . rev . 1343 (1987)
ludwig wittgenstein , philosophische untersuchungen (1953)
Steven J. Wolhandler, Voluntary Active Euthanasia for the Terminally Ill and
the Constitutional Right to Privacy, 69 cornell l . rev . 363 (1984)
Kam C. Wong, Police Powers and Control in the People’s Republic of China:
The History of Shoushen, 10 colum . j . asian l . 367 (1996)
Ledger Wood, Responsibility and Punishment, 28 am . inst . crim . l . & crimi -
nology 630 (1938)
andrew wright , gwyneth boswell , & martin davies , contemporary proba -
tion practice (1993)
Andrew J. Wu, From Video Games to Artificial Intelligence: Assigning Copy-
right Ownership to Works Generated by Increasingly Sophisticated Com-
puter Programs, 25 aipla q . j . 131 (1997)
reuven yaron , the laws of eshnunna (2d ed. 1988)
masoud yazdani & ajit narayanan , artificial intelligence : human effects
(1985)
peter young, punishment, money and the legal order: an analysis of the emer-
gence of monetary sanctions with special reference to scotland (1987)
Rachel S. Zahniser, Morally and Legally: A Parent’s Duty to Prevent the
Abuse of a Child as Defined by Lane v. Commonwealth, 86 ky . l . j . 1209
(1998)
reinhard zimmermann , the law of obligations — roman foundations of
the civilian tradition (1996)
Franklin E. Zimring, The Executioner’s Dissonant Song: On Capital Punish-
ment and American Legal Values, in the killing state — capital punish -
ment in law , politics , and culture 137 (Austin Sarat ed., 1999)
Cases
Abrams v. United States, 250 U.S. 616, 63 L.Ed. 1173, 40 S.Ct. 17 (1919)
Adams v. State, 8 Md.App. 684, 262 A.2d 69 (1970)
Alford v. State, 866 S.W.2d 619 (Tex.Crim.App.1993)
Allday, (1837) 8 Car. & P. 136, 173 E.R. 431
Almon, (1770) 5 Burr. 2686, 98 E.R. 411
Anderson v. State, 66 Okl.Cr. 291, 91 P.2d 794 (1939)
Anderson, [1966] 2 Q.B. 110, [1966] 2 All E.R. 644, [1966] 2 W.L.R. 1195, 50
Cr. App. Rep. 216, 130 J.P. 318
Andrews v. People, 800 P.2d 607 (Colo.1990)

Ann v. State, 30 Tenn. 159, 11 Hum. 159 (1850)
Arp v. State, 97 Ala. 5, 12 So. 301 (1893)
Axtell, (1660) 84 E.R. 1060
B. v. Director of Public Prosecutions, [2000] 2 A.C. 428, [2000] 1 All E.R. 833,
[2000] 2 W.L.R. 452, [2000] 2 Cr. App. Rep. 65, [2000] Crim. L.R. 403
Bailey, (1818) Russ. & Ry. 341, 168 E.R. 835
Barnes v. State, 19 Conn. 398 (1849)
Barnfather v. Islington London Borough Council, [2003] E.W.H.C. 418
(Admin), [2003] 1 W.L.R. 2318, [2003] E.L.R. 263
Bateman, [1925] All E.R. Rep. 45, 94 L.J.K.B. 791, 133 L.T. 730, 89 J.P. 162, 41
T.L.R. 557, 69 Sol. Jo. 622, 28 Cox. C.C. 33, 19 Cr. App. Rep. 8
Batson v. State, 113 Nev. 669, 941 P.2d 478 (1997)
Beason v. State, 96 Miss. 165, 50 So. 488 (1909)
Benge, (1865) 4 F. & F. 504, 176 E.R. 665
Birmingham, &c., Railway Co., (1842) 3 Q. B. 223, 114 E.R. 492
Birney v. State, 8 Ohio Rep. 230 (1837)
Blake v. United States, 407 F.2d 908 (5th Cir.1969)
Blaker v. Tillstone, [1894] 1 Q.B. 345
Bolden v. State, 171 S.W.3d 785 (2005)
Bonder v. State, 752 A.2d 1169 (Del.2000)
Boson v. Sandford, (1690) 2 Salkeld 440, 91 E.R. 382
Boushea v. United States, 173 F.2d 131, 134 (8th Cir. 1949)
Bradley v. State, 102 Tex.Crim.R. 41, 277 S.W. 147 (1926)
Bratty v. Attorney-General for Northern Ireland, [1963] A.C. 386, 409, [1961]
3 All E.R. 523, [1961] 3 W.L.R. 965, 46 Cr. App. Rep 1
Brett v. Rigden, (1568) 1 Plowd. 340, 75 E.R. 516
Burnett, (1815) 4 M. & S. 272, 105 E.R. 835
C, [2007] E.W.C.A. Crim. 1862, [2007] All E.R. (D) 91
Caldwell, [1982] A.C. 341, [1981] 1 All E.R. 961, [1981] 2 W.L.R. 509, 73 Cr.
App. Rep. 13, 145 J.P. 211
Calley v. Callaway, 519 F.2d 184 (5th Cir.1975)
Campbell v. Wood, 18 F.3d 662 (9th Cir. 1994)
Carter v. State, 376 P.2d 351 (Okl.Crim.App.1962)
Carter v. United States, 530 U.S. 255, 120 S.Ct. 2159, 147 L.Ed.2d 203 (2000)
Chance v. State, 685 A.2d 351 (Del.1996)
Cheek v. United States, 498 U.S. 192, 111 S.Ct. 604, 112 L.Ed.2d 617 (1991)
Childs v. State, 109 Nev. 1050, 864 P.2d 277 (1993)
Chisholm v. Doulton, (1889) 22 Q.B.D. 736
Chrystal v. Commonwealth, 72 Ky. 669, 9 Bush. 669 (1873)
City of Chicago v. Mayer, 56 Ill.2d 366, 308 N.E.2d 601 (1974)
Coal & C.R. v. Conley, 67 W.Va. 129, 67 S.E. 613 (1910)
Commonwealth v. Boynton, 84 Mass. 160, 2 Allen 160 (1861)
Commonwealth v. Fortner L.P. Gas Co., 610 S.W.2d 941 (Ky.App.1980)
Commonwealth v. French, 531 Pa. 42, 611 A.2d 175 (1992)

Commonwealth v. Goodman, 97 Mass. 117 (1867)
Commonwealth v. Green, 477 Pa. 170, 383 A.2d 877 (1978)
Commonwealth v. Herd, 413 Mass. 834, 604 N.E.2d 1294 (1992)
Commonwealth v. Hill, 11 Mass. 136 (1814)
Commonwealth v. Johnson, 412 Mass. 368, 589 N.E.2d 311 (1992)
Commonwealth v. Leno, 415 Mass. 835, 616 N.E.2d 453 (1993)
Commonwealth v. Lindsey, 396 Mass. 840, 489 N.E.2d 666 (1986)
Commonwealth v. McIlwain School Bus Lines, Inc., 283 Pa.Super. 1, 423
A.2d 413 (1980)
Commonwealth v. Mead, 92 Mass. 398 (1865)
Commonwealth v. Monico, 373 Mass. 298, 366 N.E.2d 1241 (1977)
Commonwealth v. New York Cent. & H. River R. Co., 206 Mass. 417, 92 N.E.
766 (1910)
Commonwealth v. Perl, 50 Mass.App.Ct. 445, 737 N.E.2d 937 (2000)
Commonwealth v. Pierce, 138 Mass. 165 (1884)
Commonwealth v. Proprietors of New Bedford Bridge, 68 Mass. 339 (1854)
Commonwealth v. Shumway, 72 Va.Cir. 481 (2007)
Commonwealth v. Thompson, 6 Mass. 134, 6 Tyng 134 (1809)
Commonwealth v. Walensky, 316 Mass. 383, 55 N.E.2d 902 (1944)
Commonwealth v. Weaver, 400 Mass. 612, 511 N.E.2d 545 (1987)
Cox v. State, 305 Ark. 244, 808 S.W.2d 306 (1991)
Crawshaw, (1860) Bell. 303, 169 E.R. 1271
Cutter v. State, 36 N.J.L. 125 (1873)
Da Silva, [2006] E.W.C.A. Crim. 1654 , [2006] 4 All E.R. 900, [2006] 2 Cr.
App. Rep. 517
Dalloway, (1847) 2 Cox C.C. 273
Daniel v. State, 187 Ga. 411, 1 S.E.2d 6 (1939)
Director of Public Prosecutions v. Kent and Sussex Contractors Ltd., [1944]
K.B. 146, [1944] 1 All E.R. 119
Dixon, (1814) 3 M. & S. 11, 105 E.R. 516
Dodd, (1736) Sess. Cas. 135, 93 E.R. 136
Dotson v. State, 6 Cold. 545 (1869)
Driver v. State, 2011 Tex. Crim. App. Lexis 4413 (2011)
Duckett v. State, 966 P.2d 941 (Wyo.1998)
Dudley and Stephens, [1884] 14 Q.B. D. 273
Dugdale, (1853) 1 El. & Bl. 435, 118 E.R. 499
Dusenbery v. Commonwealth, 263 S.E.2d 392 (Va. 1980)
Dutton v. State, 123 Md. 373, 91 A. 417 (1914)
Dyke v. Gower, [1892] 1 Q.B. 220
Elk v. United States, 177 U.S. 529, 20 S.Ct. 729, 44 L.Ed. 874 (1900)
English, [1999] A.C. 1, [1997] 4 All E.R. 545, [1997] 3 W.L.R. 959, [1998] 1 Cr.
App. Rep. 261, [1998] Crim. L.R. 48, 162 J.P. 1
Esop, (1836) 7 Car. & P. 456, 173 E.R. 203
Evans v. Bartlam, [1937] A.C. 473, 479, [1937] 2 All E.R. 646
Evans v. State, 322 Md. 24, 585 A.2d 204 (1991)
Fain v. Commonwealth, 78 Ky. 183, 39 Am.Rep. 213 (1879)
Farmer v. People, 77 Ill. 322 (1875)
Finney, (1874) 12 Cox C.C. 625
Firth, (1990) 91 Cr. App. Rep. 217, 154 J.P. 576, [1990] Crim. L.R. 326
Fitzpatrick v. Kelly, [1873] 8 Q.B. 337
Fitzpatrick, [1977] N.I. 20
Forbes, (1835) 7 Car. & P. 224, 173 E.R. 99
Frey v. United States, 708 So.2d 918 (Fla.1998)
G., [2003] U.K.H.L. 50, [2004] 1 A.C. 1034, [2003] 3 W.L.R. 1060, [2003] 4 All
E.R. 765, [2004] 1 Cr. App. Rep. 21, (2003) 167 J.P. 621, [2004] Crim. L. R.
369
G., [2008] U.K.H.L. 37, [2009] A.C. 92
Gammon (Hong Kong) Ltd. v. Attorney-General of Hong Kong, [1985] 1 A.C.
1, [1984] 2 All E.R. 503, [1984] 3 W.L.R. 437, 80 Cr. App. Rep. 194, 26
Build L.R. 159
Gardiner, [1994] Crim. L.R. 455
Godfrey v. State, 31 Ala. 323 (1858)
Government of the Virgin Islands v. Smith, 278 F.2d 169 (3rd Cir.1960)
Granite Construction Co. v. Superior Court, 149 Cal.App.3d 465, 197 Cal.
Rptr. 3 (1983)
Gray v. Lucas, 710 F.2d 1048 (5th Cir. 1983)
Great Broughton (Inhabitants), (1771) 5 Burr. 2700, 98 E.R. 418
Gregg v. Georgia, 428 U.S. 153, S.Ct. 2909, 49 L.Ed.2d 859 (1979)
Grout, (1834) 6 Car. & P. 629, 172 E.R. 1394
Hall v. Brooklands Auto Racing Club, [1932] All E.R. 208, [1933] 1 K.B. 205,
101 L.J.K.B. 679, 147 L.T. 404, 48 T.L.R. 546
Hardcastle v. Bielby, [1892] 1 Q.B. 709
Hartson v. People, 125 Colo. 1, 240 P.2d 907 (1951)
Hasan, [2005] U.K.H.L. 22, [2005] 4 All E.R. 685, [2005] 2 Cr. App. Rep. 314,
[2006] Crim. L.R. 142, [2005] All E.R. (D) 299
Heilman v. Commonwealth, 84 Ky. 457, 1 S.W. 731 (1886)
Henderson v. Kibbe, 431 U.S. 145, 97 S.Ct. 1730, 52 L.Ed.2d 203 (1977)
Henderson v. State, 11 Ala.App. 37, 65 So. 721 (1914)
Hentzner v. State, 613 P.2d 821 (Alaska 1980)
Hern v. Nichols, (1708) 1 Salkeld 289, 91 E.R. 256
Hobbs v. Winchester Corporation, [1910] 2 K.B. 471
Holbrook, (1878) 4 Q.B.D. 42
Howard v. State, 73 Ga.App. 265, 36 S.E.2d 161 (1945)
Huggins, (1730) 2 Strange 882, 93 E.R. 915
Hughes v. Commonwealth, 19 Ky.L.R. 497, 41 S.W. 294 (1897)
Hull, (1664) Kel. 40, 84 E.R. 1072
Humphrey v. Commonwealth, 37 Va.App. 36, 553 S.E.2d 546 (2001)
Hunt v. Nuth, 57 F.3d 1327 (4th Cir. 1995)

Hunt v. State, 753 So.2d 609 (Fla.App.2000)
Hunter v. State, 30 Tenn. 160, 1 Head 160 (1858)
I.C.R. Haulage Ltd., [1944] K.B. 551, [1944] 1 All E.R. 691
Ingram v. United States, 592 A.2d 992 (D.C.App.1991)
Johnson v. State, 142 Ala. 70 (1904)
Jones v. Hart, (1699) 2 Salkeld 441, 91 E.R. 382
Jurco v. State, 825 P.2d 909 (Alaska App.1992)
K., [2001] U.K.H.L. 41, [2002] 1 A.C. 462
Kimoktoak v. State, 584 P.2d 25 (Alaska 1978)
Kingston v. Booth, (1685) Skinner 228, 90 E.R. 105
Knight, (1828) 1 L.C.C. 168, 168 E.R. 1000
Kumar, [2004] E.W.C.A. Crim. 3207, [2005] 1 Cr. App. Rep. 566, [2005] Crim.
L.R. 470
Laaman v. Helgemoe, 437 F.Supp. 269 (1977)
Lambert v. California, 355 U.S. 225, 78 S.Ct. 240, 2 L.Ed.2d 228 (1957)
Lambert v. State, 374 P.2d 783 (Okla.Crim.App.1962)
Lane v. Commonwealth, 956 S.W.2d 874 (Ky.1997)
Langforth Bridge, (1635) Cro. Car. 365, 79 E.R. 919
Larsonneur, (1933) 24 Cr. App. R. 74, 97 J.P. 206, 149 L.T. 542
Lawson, [1986] V.R. 515
Leach, [1937] 1 All E.R. 319
Lee v. State, 41 Tenn. 62, 1 Cold. 62 (1860)
Leet v. State, 595 So.2d 959 (1991)
Lennard’s Carrying Co. Ltd. v. Asiatic Petroleum Co. Ltd., [1915] A.C. 705
In re Leroy, 285 Md. 508, 403 A.2d 1226 (1979)
Lester v. State, 212 Tenn. 338, 370 S.W.2d 405 (1963)
Levett, (1638) Cro. Car. 538
lifton (Inhabitants), (1794) 5 T.R. 498, 101 E.R. 280
Liverpool (Mayor), (1802) 3 East 82, 102 E.R. 529
Long v. Commonwealth, 23 Va.App. 537, 478 S.E.2d 324 (1996)
Long v. State, 44 Del. 262, 65 A.2d 489 (1949)
Longbottom, (1849) 3 Cox C. C. 439
Lutwin v. State, 97 N.J.L. 67, 117 A. 164 (1922)
Manser, (1584) 2 Co. Rep. 3, 76 E.R. 392
Marshall, (1830) 1 Lewin 76, 168 E.R. 965
Martin v. State, 90 Ala. 602, 8 So. 858 (1891)
Mason v. State, 603 P.2d 1146 (Okl.Crim.App.1979)
Matudi, [2004] E.W.C.A. Crim. 697
Mavji, [1987] 2 All E.R. 758, [1987] 1 W.L.R. 1388, [1986] S.T.C. 508, Cr. App.
Rep. 31, [1987] Crim. L.R. 39
Maxey v. United States, 30 App. D.C. 63, 80 (App. D.C. 1907)
McClain v. State, 678 N.E.2d 104 (Ind.1997)
McGrowther, (1746) 18 How. St. Tr. 394
McMillan v. City of Jackson, 701 So.2d 1105 (Miss.1997)

McNeil v. United States, 933 A.2d 354 (2007)
Meade, [1909] 1 K.B. 895
Meakin, (1836) 7 Car. & P. 297, 173 E.R. 131
Mendez v. State, 575 S.W.2d 36 (Tex.Crim.App.1979)
Michael, (1840) 2 Mood. 120, 169 E.R. 48
Middleton v. Fowler, (1699) 1 Salkeld 282, 91 E.R. 247
Mildmay, (1584) 1 Co. Rep. 175a, 76 E.R. 379
Miller v. State, 3 Ohio St. Rep. 475 (1854)
Minor v. State, 326 Md. 436, 605 A.2d 138 (1992)
Mitchell v. State, 114 Nev. 1417, 971 P.2d 813 (1998)
M’Naghten, (1843) 10 Cl. & Fin. 200, 8 E.R. 718
Montgomery v. Commonwealth, 189 Ky. 306, 224 S.W. 878 (1920)
Moore v. State, 25 Okl.Crim. 118, 218 P. 1102 (1923)
Mouse, (1608) 12 Co. Rep. 63, 77 E.R. 1341
Myers v. State, 1 Conn. 502 (1816)
Nelson v. State, 597 P.2d 977 (Alaska 1979)
New York & G.L.R. Co. v. State, 50 N.J.L. 303, 13 A. 1 (1888)
New York Cent. & H.R.R. v. United States, 212 U.S. 481, 29 S.Ct. 304, 53
L.Ed. 613 (1909)
Nutt, (1728) 1 Barn. K.B. 306, 94 E.R. 208
O’Flaherty, [2004] E.W.C.A. Crim. 526, [2004] 2 Cr. App. Rep. 315
Oxford, (1840) 9 Car. & P. 525, 173 E.R. 941
Parish, (1837) 8 Car. & P. 94, 173 E.R. 413
Parnell v. State, 912 S.W.2d 422, 424 (Ark. 1996)
Pearson, (1835) 2 Lewin 144, 168 E.R. 1108
Peebles v. State, 101 Ga. 585, 28 S.E. 920 (1897)
People v. Bailey, 451 Mich. 657, 549 N.W.2d 325 (1996)
People v. Brubaker, 53 Cal.2d 37, 346 P.2d 8 (1959)
People v. Cabaltero, 31 Cal.App.2d 52, 87 P.2d 364 (1939)
People v. Cherry, 307 N.Y. 308, 121 N.E.2d 238 (1954)
People v. Clark, 8 N.Y.Cr. 169, 14 N.Y.S. 642 (1891)
People v. Cooper, 194 Ill.2d 419, 252 Ill.Dec. 458, 743 N.E.2d 32 (2000)
People v. Craig, 78 N.Y.2d 616, 578 N.Y.S.2d 471, 585 N.E.2d 783 (1991)
People v. Daugherty, 40 Cal.2d 876, 256 P.2d 911 (1953)
People v. Davis, 33 N.Y.2d 221, 351 N.Y.S.2d 663, 306 N.E.2d 787 (1973)
People v. Decina, 2 N.Y.2d 133, 157 N.Y.S.2d 558, 138 N.E.2d 799 (1956)
People v. Disimone, 251 Mich.App. 605, 650 N.W.2d 436 (2002)
People v. Ferguson, 134 Cal.App. 41, 24 P.2d 965 (1933)
People v. Freeman, 61 Cal.App.2d 110, 142 P.2d 435 (1943)
People v. Handy, 198 Colo. 556, 603 P.2d 941 (1979)
People v. Haney, 30 N.Y.2d 328, 333 N.Y.S.2d 403, 284 N.E.2d 564 (1972)
People v. Harris, 29 Cal. 678 (1866)
People v. Heitzman, 9 Cal.4th 189, 37 Cal.Rptr.2d 236, 886 P.2d 1229 (1994)
People v. Henry, 239 Mich.App. 140, 607 N.W.2d 767 (1999)

People v. Higgins, 5 N.Y.2d 607, 186 N.Y.S.2d 623, 159 N.E.2d 179 (1959)
People v. Howk, 56 Cal.2d 687, 16 Cal.Rptr. 370, 365 P.2d 426 (1961)
People v. Kemp, 150 Cal.App.2d 654, 310 P.2d 680 (1957)
People v. Kessler, 57 Ill.2d 493, 315 N.E.2d 29 (1974)
People v. Kirst, 168 N.Y. 19, 60 N.E. 1057 (1901)
People v. Larkins, 2010 Mich. App. Lexis 1891 (2010)
People v. Leonardi, 143 N.Y. 360, 38 N.E. 372 (1894)
People v. Lisnow, 88 Cal.App.3d Supp. 21, 151 Cal.Rptr. 621 (1978)
People v. Little, 41 Cal.App.2d 797, 107 P.2d 634 (1941)
People v. Marshall, 362 Mich. 170, 106 N.W.2d 842 (1961)
People v. Merhige, 212 Mich. 601, 180 N.W. 418 (1920)
People v. Michalow, 229 N.Y. 325, 128 N.E. 228 (1920)
People v. Minifie, 13 Cal.4th 1055, 56 Cal.Rptr.2d 133, 920 P.2d 1337 (1996)
People v. Monks, 133 Cal. App. 440 (Cal. Dist. Ct. App. 1933)
People v. Mutchler, 140 N.E. 820, 823 (Ill. 1923)
People v. Newton, 8 Cal.App.3d 359, 87 Cal.Rptr. 394 (1970)
People v. Pantano, 239 N.Y. 416, 146 N.E. 646 (1925)
People v. Prettyman, 14 Cal.4th 248, 58 Cal.Rptr.2d 827, 926 P.2d 1013 (1996)
People v. Richards, 269 Cal.App.2d 768, 75 Cal.Rptr. 597 (1969)
People v. Sakow, 45 N.Y.2d 131, 408 N.Y.S.2d 27, 379 N.E.2d 1157 (1978)
People v. Smith, 57 Cal. App. 4th 1470, 67 Cal. Rptr. 2d 604 (1997)
People v. Sommers, 200 P.3d 1089 (2008)
People v. Townsend, 214 Mich. 267, 183 N.W. 177 (1921)
People v. Vogel, 46 Cal.2d 798, 299 P.2d 850 (1956)
People v. Weiss, 256 App.Div. 162, 9 N.Y.S.2d 1 (1939)
People v. Whipple, 100 Cal.App. 261, 279 P. 1008 (1929)
People v. Williams, 56 Ill.App.2d 159, 205 N.E.2d 749 (1965)
People v. Wilson, 66 Cal.2d 749, 59 Cal.Rptr. 156, 427 P.2d 820 (1967)
People v. Young, 11 N.Y.2d 274, 229 N.Y.S.2d 1, 183 N.E.2d 319 (1962)
Pierson v. State, 956 P.2d 1119 (Wyo.1998)
Pigman v. State, 14 Ohio 555 (1846)
Polston v. State, 685 P.2d 1 (Wyo.1984)
Pope v. United States, 372 F.2d 710 (8th Cir.1970)
Powell v. Texas, 392 U.S. 514, 88 S.Ct. 2145, 20 L.Ed.2d 1254 (1968)
Price v. State, 50 Tex.Crim.R. 71, 94 S.W. 901 (1906)
Proctor v. State, 15 Okl.Cr. 338, 176 P. 771 (1918)
Provenzano v. Moore, 744 So.2d 413 (Fla. 1999)
Provincial Motor Cab Company Ltd. v. Dunning, [1909] 2 K.B. 599
Pugliese v. Commonwealth, 16 Va.App. 82, 428 S.E.2d 16 (1993)
Quick, [1973] Q.B. 910, [1973] 3 All E.R. 347, [1973] 3 W.L.R. 26, 57 Cr. App.
Rep. 722, 137 J.P. 763
R.I. Recreation Center v. Aetna Cas. & Surety Co., 177 F.2d 603 (1st Cir.1949)
Rangel v. State, 2009 Tex.App. 1555 (2009)
Ratzlaf v. United States, 510 U.S. 135, 114 S.Ct. 655, 126 L.Ed.2d 615 (1994)
Read v. People, 119 Colo. 506, 205 P.2d 233 (1949)
Redmond v. State, 36 Ark. 58 (1880)
Reed v. State, 693 N.E.2d 988 (Ind.App.1998)
Reniger v. Fogossa, (1551) 1 Plowd. 1, 75 E.R. 1
Rice v. State, 8 Mo. 403 (1844)
Richards, [2004] E.W.C.A. Crim. 192
Richardson v. State, 697 N.E.2d 462 (Ind.1998)
Ricketts v. State, 291 Md. 701, 436 A.2d 906 (1981)
Roberts v. People, 19 Mich. 401 (1870)
Robinson v. California, 370 U.S. 660, 82 S.Ct. 1417, 8 L.Ed.2d 758 (1962)
Rollins v. State, 2009 Ark. 484, 347 S.W.3d 20 (2009)
Roy v. United States, 652 A.2d 1098 (D.C.App.1995)
Saik, [2006] U.K.H.L. 18, [2007] 1 A.C. 18
Saintiff, (1705) 6 Mod. 255, 87 E.R. 1002
Sam v. Commonwealth, 13 Va.App. 312, 411 S.E.2d 832 (1991)
Sanders v. State, 466 N.E.2d 424 (Ind.1984)
Scales v. United States, 367 U.S. 203, 81 S.Ct. 1469, 6 L.Ed.2d 782 (1961)
Schmidt v. United States, 133 F. 257 (9th Cir.1904)
Schuster v. State, 48 Ala. 199 (1872)
Scott v. State, 71 Tex.Crim.R. 41, 158 S.W. 814 (1913)
Seaboard Offshore Ltd. v. Secretary of State for Transport, [1994] 2 All E.R.
99, [1994] 1 W.L.R. 541, [1994] 1 Lloyd’s Rep. 593
Seaman v. Browning, (1589) 4 Leonard 123, 74 E.R. 771
Severn and Wye Railway Co., (1819) 2 B. & Ald. 646, 106 E.R. 501
Ex parte Smith, 135 Mo. 223, 36 S.W. 628 (1896)
Smith v. California, 361 U.S. 147, 80 S.Ct. 215, 4 L.Ed.2d 205 (1959)
Smith v. State, 83 Ala. 26, 3 So. 551 (1888)
Spiers & Pond v. Bennett, [1896] 2 Q.B. 65
Squire v. State, 46 Ind. 459 (1874)
State Philbrick, 402 A.2d 59 (Me.1979)
State v. Aaron, 4 N.J.L. 269 (1818)
State v. Anderson, 141 Wash.2d 357, 5 P.3d 1247 (2000)
State v. Anthuber, 201 Wis.2d 512, 549 N.W.2d 477 (App.1996)
State v. Asher, 50 Ark. 427, 8 S.W. 177 (1888)
State v. Audette, 149 Vt. 218, 543 A.2d 1315 (1988)
State v. Ayer, 136 N.H. 191, 612 A.2d 923 (1992)
State v. Barker, 128 W.Va. 744, 38 S.E.2d 346 (1946)
State v. Barrett, 768 A.2d 929 (R.I.2001)
State v. Blakely, 399 N.W.2d 317 (S.D.1987)
State v. Bono, 128 N.J.Super. 254, 319 A.2d 762 (1974)
State v. Bowen, 118 Kan. 31, 234 P. 46 (1925)
State v. Brosnan, 221 Conn. 788, 608 A.2d 49 (1992)
State v. Brown, 389 So.2d 48 (La.1980)
State v. Bunkley, 202 Conn. 629, 522 A.2d 795 (1987)

State v. Burrell, 135 N.H. 715, 609 A.2d 751 (1992)
State v. Cain, 9 W. Va. 559 (1874)
State v. Cameron, 104 N.J. 42, 514 A.2d 1302 (1986)
State v. Campbell, 536 P.2d 105 (Alaska 1975)
State v. Carrasco, 122 N.M. 554, 928 P.2d 939 (1996)
State v. Case, 672 A.2d 586 (Me.1996)
State v. Caswell, 771 A.2d 375 (Me.2001)
State v. Champa, 494 A.2d 102 (R.I.1985)
State v. Chicago, M. & St.P.R. Co., 130 Minn. 144, 153 N.W. 320 (1915)
State v. Clottu, 33 Ind. 409 (1870)
State v. Coffin, 128 N.M. 192, 991 P.2d 477 (1999)
State v. Cram, 157 Vt. 466, 600 A.2d 733 (1991)
State v. Crocker, 431 A.2d 1323 (Me.1981)
State v. Crocker, 506 A.2d 209 (Me.1986)
State v. Cude, 14 Utah 2d 287, 383 P.2d 399 (1963)
State v. Curry, 45 Ohio St.3d 109, 543 N.E.2d 1228 (1989)
State v. Daniels, 236 La. 998, 109 So.2d 896 (1958)
State v. Dansinger, 521 A.2d 685 (Me.1987)
State v. Daoud, 141 N.H. 142, 679 A.2d 577 (1996)
State v. Dillon, 93 Idaho 698, 471 P.2d 553 (1970)
State v. Dubina, 164 Conn. 95, 318 A.2d 95 (1972)
State v. Ehlers, 98 N.J.L. 263, 119 A. 15 (1922)
State v. Ellis, 232 Or. 70, 374 P.2d 461 (1962)
State v. Elsea, 251 S.W.2d 650 (Mo.1952)
State v. Etzweiler, 125 N.H. 57, 480 A.2d 870 (1984)
State v. Evans, 134 N.H. 378, 594 A.2d 154 (1991)
State v. Farley, 225 Kan. 127, 587 P.2d 337 (1978)
State v. Fee, 126 N.H. 78, 489 A.2d 606 (1985)
State v. Finnell, 101 N.M. 732, 688 P.2d 769 (1984)
State v. Fletcher, 322 N.C. 415, 368 S.E.2d 633 (1988)
State v. Follin, 263 Kan. 28, 947 P.2d 8 (1997)
State v. Foster, 202 Conn. 520, 522 A.2d 277 (1987)
State v. Foster, 91 Wash.2d 466, 589 P.2d 789 (1979)
State v. Gallagher, 191 Conn. 433, 465 A.2d 323 (1983)
State v. Gartland, 304 Mo. 87, 263 S.W. 165 (1924)
State v. Garza, 259 Kan. 826, 916 P.2d 9 (1996)
State v. George, 20 Del. 57, 54 A. 745 (1902)
State v. Gish, 17 Idaho 341, 393 P.2d 342 (1964)
State v. Goodall, 407 A.2d 268 (Me.1979)
State v. Goodenow, 65 Me. 30 (1876)
State v. Gray, 221 Conn. 713, 607 A.2d 391 (1992)
State v. Great Works Mill. & Mfg. Co., 20 Me. 41, 37 Am.Dec.38 (1841)
State v. Hadley, 65 Utah 109, 234 P. 940 (1925)
State v. Harris, 222 N.W.2d 462 (Iowa 1974)

State v. Hastings, 118 Idaho 854, 801 P.2d 563 (1990)
State v. Havican, 213 Conn. 593, 569 A.2d 1089 (1990)
State v. Herro, 120 Ariz. 604, 587 P.2d 1181 (1978)
State v. Hinkle, 200 W.Va. 280, 489 S.E.2d 257 (1996)
State v. Hobbs, 252 Iowa 432, 107 N.W.2d 238 (1961)
State v. Hooker, 17 Vt. 658 (1845)
State v. Hopkins, 147 Wash. 198, 265 P. 481 (1928)
State v. Howley, 128 Idaho 874, 920 P.2d 391 (1996)
State v. I. & M. Amusements, Inc., 10 Ohio App.2d 153, 226 N.E.2d 567 (1966)
State v. J.P.S., 135 Wash.2d 34, 954 P.2d 894 (1998)
State v. Jackson, 137 Wash.2d 712, 976 P.2d 1229 (1999)
State v. Jackson, 346 Mo. 474, 142 S.W.2d 45 (1940)
State v. Jacobs, 371 So.2d 801 (La.1979)
State v. Jenner, 451 N.W.2d 710 (S.D.1990)
State v. Johnson, 233 Wis. 668, 290 N.W. 159 (1940)
State v. Kaiser, 260 Kan. 235, 918 P.2d 629 (1996)
State v. Kee, 398 A.2d 384 (Me.1979)
State v. Labato, 7 N.J. 137, 80 A.2d 617 (1951)
State v. Lawrence, 97 N.C. 492, 2 S.E. 367 (1887)
State v. Linscott, 520 A.2d 1067 (Me.1987)
State v. Lockhart, 208 W.Va. 622, 542 S.E.2d 443 (2000)
State v. Lucas, 55 Iowa 321, 7 N.W. 583 (1880)
State v. Marley, 54 Haw. 450, 509 P.2d 1095 (1973)
State v. Martin, 119 N.J. 2, 573 A.2d 1359 (1990)
State v. McDowell, 312 N.W.2d 301 (N.D. 1981)
State v. Mendoza, 709 A.2d 1030 (R.I.1998)
State v. Mishne, 427 A.2d 450 (Me.1981)
State v. Molin, 288 N.W.2d 232 (Minn.1979)
State v. Moore, 158 N.J. 292, 729 A.2d 1021 (1999)
State v. Murphy, 674 P.2d 1220 (Utah.1983)
State v. Nargashian, 26 R.I. 299, 58 A. 953 (1904)
State v. Nelson, 329 N.W.2d 643 (Iowa 1983)
State v. Neuzil, 589 N.W.2d 708 (Iowa 1999)
State v. Nickelson, 45 La.Ann. 1172, 14 So. 134 (1893)
State v. Pereira, 72 Conn. App. 545, 805 A.2d 787 (2002)
State v. Pincus, 41 N.J.Super. 454, 125 A.2d 420 (1956)
State v. Reed, 205 Neb. 45, 286 N.W.2d 111 (1979)
State v. Reese, 272 N.W.2d 863 (Iowa 1978)
State v. Robinson, 132 Ohio App.3d 830, 726 N.E.2d 581 (1999)
State v. Robinson, 20 W.Va. 713, 43 Am.Rep. 799 (1882)
State v. Rocheville, 310 S.C. 20, 425 S.E.2d 32 (1993)
State v. Rocker, 52 Haw. 336, 475 P.2d 684 (1970)
State v. Runkles, 605 A.2d 111, 121 (Md. 1992)
State v. Sargent, 156 Vt. 463, 594 A.2d 401 (1991)

State v. Sasse, 6 S.D. 212, 60 N.W. 853 (1894)
State v. Sawyer, 95 Conn. 34, 110 A. 461 (1920)
State v. Schulz, 55 Ia. 628 (1881)
State v. Sexton, 160 N.J. 93, 733 A.2d 1125 (1999)
State v. Sheedy, 125 N.H. 108, 480 A.2d 887 (1984)
State v. Silva-Baltazar, 125 Wash.2d 472, 886 P.2d 138 (1994)
State v. Silveira, 198 Conn. 454, 503 A.2d 599 (1986)
State v. Smith, 170 Wis.2d 701, 490 N.W.2d 40 (App.1992)
State v. Smith, 219 N.W.2d 655 (Iowa 1974)
State v. Smith, 260 Or. 349, 490 P.2d 1262 (1971)
State v. Stepniewski, 105 Wis.2d 261, 314 N.W.2d 98 (1982)
State v. Stewart, 624 N.W.2d 585 (Minn.2001)
State v. Stoehr, 134 Wis.2d 66, 396 N.W.2d 177 (1986)
State v. Striggles, 202 Iowa 1318, 210 N.W. 137 (1926)
State v. Strong, 294 N.W.2d 319 (Minn.1980)
State v. Thomas, 619 S.W.2d 513, 514 (Tenn. 1981)
State v. Torphy, 78 Mo.App. 206 (1899)
State v. Torres, 495 N.W.2d 678 (1993)
State v. Totman, 80 Mo.App. 125 (1899)
State v. VanTreese, 198 Iowa 984, 200 N.W. 570 (1924)
State v. Warshow, 138 Vt. 22, 410 A.2d 1000 (1979)
State v. Welsh, 8 Wash.App. 719, 508 P.2d 1041 (1973)
State v. Wenger, 58 Ohio St.2d 336, 390 N.E.2d 801 (1979)
State v. Whitman, 116 Fla. 196, 156 So. 705 (1934)
State v. Whitoomb, 52 Iowa 85, 2 N.W. 970 (1879)
State v. Wilchinski, 242 Conn. 211, 700 A.2d 1 (1997)
State v. Wilson, 267 Kan. 550, 987 P.2d 1060 (1999)
State v. Wyatt, 198 W.Va. 530, 482 S.E.2d 147 (1996)
Stein v. State, 37 Ala. 123 (1861)
Stephens, [1866] 1 Q.B. 702
Stratford-upon-Avon Corporation, (1811) 14 East 348, 104 E.R. 636
Studstill v. State, 7 Ga. 2 (1849)
Sweet v. Parsley, [1970] A.C. 132, [1969] 1 All E.R. 347, [1969] 2 W.L.R. 470,
133 J.P. 188, 53 Cr. App. Rep. 221, 209 E.G. 703, [1969] E.G.D. 123
Tate v. Commonwealth, 258 Ky. 685, 80 S.W.2d 817 (1935)
Taylor v. State, 158 Miss. 505, 130 So. 502 (1930)
Texaco Inc. v. Short, 454 U.S. 516, 102 S.Ct. 781, 70 L.Ed.2d 738 (1982)
Thomas, (1837) 7 Car. & P. 817, 173 E.R. 356
Thompson v. State, 44 S.W.3d 171 (Tex.App.2001)
Thompson v. United States, 348 F.Supp.2d 398 (2005)
Tift v. State, 17 Ga.App. 663, 88 S.E. 41 (1916)
Treacy v. Director of Public Prosecutions, [1971] A.C. 537, 559, [1971] 1 All
E.R. 110, [1971] 2 W.L.R. 112, 55 Cr. App. Rep. 113, 135 J.P. 112
Tully v. State, 730 P.2d 1206 (Okl.Crim.App.1986)

Turberwill v. Stamp, (1697) Skinner 681, 90 E.R. 303
In re Tyvonne, 211 Conn. 151, 558 A.2d 661 (1989)
United States v. Alaska Packers’ Association, 1 Alaska 217 (1901)
United States v. Albertini, 830 F.2d 985 (9th Cir.1987)
United States v. Allegheny Bottling Company, 695 F.Supp. 856 (1988)
United States v. Andrews, 75 F.3d 552 (9th Cir.1996)
United States v. Arthurs, 73 F.3d 444 (1st Cir.1996)
United States v. Bailey, 444 U.S. 394, 100 S.Ct. 624, 62 L.Ed.2d 575 (1980)
United States v. Bakhtiari, 913 F.2d 1053 (2nd Cir.1990)
United States v. Bryan, 483 F.2d 88, 92 (3d Cir. 1973)
United States v. Buber, 62 M.J. 476 (2006)
United States v. Calley, 48 C.M.R. 19, 22 U.S.C.M.A. 534 (1973)
United States v. Campbell, 675 F.2d 815 (6th Cir.1982)
United States v. Carter, 311 F.2d 934 (6th Cir.1963)
United States v. Chandler, 393 F.2d 920 (4th Cir.1968)
United States v. Contento-Pachon, 723 F.2d 691 (9th Cir.1984)
United States v. Currens, 290 F.2d 751 (3rd Cir.1961)
United States v. Doe, 136 F.3d 631 (9th Cir.1998)
United States v. Dominguez-Ochoa, 386 F.3d 639 (2004)
United States v. Dorrell, 758 F.2d 427 (9th Cir.1985)
United States v. Dye Construction Co., 510 F.2d 78 (10th Cir.1975)
United States v. Freeman, 25 Fed. Cas. 1208 (1827)
United States v. Freeman, 357 F.2d 606 (2nd Cir.1966)
United States v. Gomez, 81 F.3d 846 (9th Cir.1996)
United States v. Greer, 467 F.2d 1064 (7th Cir.1972)
United States v. Hanousek, 176 F.3d 1116 (9th Cir.1999)
United States v. Heredia, 483 F.3d 913 (2006)
United States v. Holmes, 26 F. Cas. 360, 1 Wall. Jr. 1 (1842)
United States v. Jewell, 532 F.2d 697 (9th Cir.1976)
United States v. John Kelso Co., 86 F. 304 (Cal.1898)
United States v. Johnson, 956 F.2d 894 (9th Cir.1992)
United States v. Kabat, 797 F.2d 580 (8th Cir.1986)
United States v. Ladish Malting Co., 135 F.3d 484 (7th Cir.1998)
United States v. LaFleur, 971 F.2d 200 (9th Cir.1991)
United States v. Lampkins, 4 U.S.C.M.A. 31, 15 C.M.R. 31 (1954)
United States v. Lee, 694 F.2d 649 (11th Cir.1983)
United States v. Mancuso, 139 F.2d 90 (3rd Cir.1943)
United States v. Maxwell, 254 F.3d 21 (1st Cir.2001)
United States v. Meyers, 906 F. Supp. 1494 (1995)
United States v. Moore, 486 F.2d 1139 (D.C.Cir.1973)
United States v. Oakland Cannabis Buyers’ Cooperative, 532 U.S. 483, 121
S.Ct. 1711, 149 L.Ed.2d 722 (2001)
United States v. Paolello, 951 F.2d 537 (3rd Cir.1991)
United States v. Pomponio, 429 U.S. 10, 97 S.Ct. 22, 50 L.Ed.2d 12 (1976)
United States v. Powell, 929 F.2d 724 (D.C.Cir.1991)
United States v. Quaintance, 471 F. Supp.2d 1153 (2006)
United States v. Ramon-Rodriguez, 492 F.3d 930 (2007)
United States v. Randall, 104 Wash.D.C.Rep. 2249 (D.C.Super.1976)
United States v. Randolph, 93 F.3d 656 (9th Cir.1996)
United States v. Robertson, 33 M.J. 832 (1991)
United States v. Ruffin, 613 F.2d 408 (2d Cir. 1979)
United States v. Shapiro, 383 F.2d 680 (7th Cir.1967)
United States v. Smith, 404 F.2d 720 (6th Cir.1968)
United States v. Spinney, 65 F.3d 231 (1st Cir.1995)
United States v. Sued-Jimenez, 275 F.3d 1 (1st Cir.2001)
United States v. Thompson-Powell Drilling Co., 196 F.Supp. 571
(N.D.Tex.1961)
United States v. Tobon-Builes, 706 F.2d 1092 (11th Cir. 1983)
United States v. Torres, 977 F.2d 321 (7th Cir.1992)
United States v. Warner, 28 Fed. Cas. 404, 6 W.L.J. 255, 4 McLean 463 (1848)
United States v. Wert-Ruiz, 228 F.3d 250 (3rd Cir.2000)
United States v. Youts, 229 F.3d 1312 (10th Cir.2000)
Vantandillo, (1815) 4 M. & S. 73, 105 E.R. 762
Vaux, (1613) 1 Blustrode 197, 80 E.R. 885
Virgin Islands v. Joyce, 210 F. App. 208 (2006)
Walter, (1799) 3 Esp. 21, 170 E.R. 524
Webb, [2006] E.W.C.A. Crim. 2496, [2007] All E.R. (D) 406
In re Welfare of C.R.M., 611 N.W.2d 802 (Minn.2000)
Wheatley v. Commonwealth, 26 Ky.L.Rep. 436, 81 S.W. 687 (1904)
Wieland v. State, 101 Md. App. 1, 643 A.2d 446 (1994)
Wilkerson v. Utah, 99 U.S. (9 Otto) 130, 25 L.Ed. 345 (1878)
Willet v. Commonwealth, 76 Ky. 230 (1877)
Williams v. State, 70 Ga.App. 10, 27 S.E.2d 109 (1943)
Williamson, (1807) 3 Car. & P. 635, 172 E.R. 579
Wilson v. State, 24 S.W. 409 (Tex.Crim.App.1893)
Wilson v. State, 777 S.W.2d 823 (Tex.App.1989)
In re Winship, 397 U.S. 358, 90 S.Ct. 1068, 25 L.Ed.2d 368 (1970)
Woodrow, (1846) 15 M. & W. 404, 153 E.R. 907
Index
Note: Page numbers in italics refer to circumstances: and duress, 147; and

tables infancy, 126; and necessity, 147;
and negligence, 85, 88–89, 89,
absolute defense model, 150 91–92, 102; and mental element
absolute responsibility model, 150 requirement, 45–49, 49, 52, 55–56;
accessoryship, 29, 41, 69 optional component, 38; and
accountability, 17–19 perpetration-through-another, 72;
act, 39–41 and probable consequence liability,
actio libera in causa, 127, 141, 145, 148 79–80; and self-defense, 141, 144;
animals, 7, 22–25 and sentencing, 156, 163, 167, 175;
Asimov, 15–17, 141, 143 and strict liability, 105, 108, 109;
attention, 2, 50, 54 structure of factual element require
awareness: and AI systems, 125; cog ment, 30–32, 31, 42–43; and sub
nition and volition, 31–33, 33, stantive immunity, 139
131–136, 154; definition of, in crim cognition, 31–33, 33, 45–48, 49, 51, 54,
inal law, 49–56, 63–64, 66–67, 61, 88, 136
98–100; and mens rea, 45–48, combat decisions, 4
49, 98, 120–121; and negligence, combinatorial explosion, 10
86–89, 89, 90–91, 94–95, 98, commonsense, 3
101–102, 118; self-awareness, 8, communication, 7, 9, 24
127–128; and will, 58, 71–72, 82. community service, 171
See also mens rea; negligence complicity, 29, 68–69, 71, 78, 99
conduct: and AI entity liability, 65;
Babbage, 2 and general defenses, 121, 129,
blindness, 52, 55–56. See also willful 147; goal driven, 8–9; and mental
blindness element requirement, 38–49, 49,
bona fide, 137, 139 55–60, 63; and negligence, 85–89,
bureaucracy, 10 89, 92–93; and perpetration-
through-another, 69, 72–73, 114;
capital punishment, 142, 157, 162, and probable consequence liability,
165–167, 169 76; and sentencing, 169; and strict
characteristica universalis, 1 liability, 105–108, 109, 118; struc
chess, 2, 4, 11, 59–60 ture of factual element requirement,
chinese room, 6–7 26–32, 31
241
index . 242
corporations: and general criminal lia in general defenses, 120, 129, 131,
bility, 21–22, 34–36; and general de 133–136, 154; and sentencing 160
fenses, 124–126, 142–144, 153, 155; fine, 26, 106, 158, 162–163, 165, 173–175
and intentional offenses, 66–69; foreseeability, 58–64, 79, 100–101, 117
and negligence, 96–97, 99; and sen formal logic, 6, 8, 29
tencing, 162–165, 167, 170–175;
and strict liability, 112, 114 general intent, 32, 44, 57
creativity, 7, 9, 20 generalization, 10–11, 84, 92, 100
general problem solver, 2, 6
de minimis, 123, 152–154 good faith. See bona fide
death penalty. See capital punishment
Deep Blue, 11 heuristics, 3
deontological morality, 18 Hobbes, 1
Descartes, 1 Hollywood, 15
deterrence, 76, 81, 157–159, 161, homicide, 25, 32; and general defenses,
166–168, 170, 172, 174 121, 131, 147, 154; intentional, 38,
discipline, 150, 167 42–44, 48, 75–76, 78–80, 82–83;
dreaming, 134 negligent, 86, 94, 96, 98, 100–103;
drone, 4, 104, 143, 151–152 and sentencing, 166; strict liability,
duress, 120, 123, 140, 147–149, 155 117–119
dynamic modification, 10 humanoid, 14, 25
human-robot coexistence, 14–15, 17, 22
evil, 25–26, 33, 44, 56, 64–65, 68; and
necessity, 145–147; and negligence, imminent danger, 141–142, 145, 148
96 imprisonment, 26, 142, 156, 162–165,
exemptions, 122–123 167–169, 171, 173, 175–176
ex officio, 138 inaction, 16–17, 28, 39–41, 87
expert systems, 3–4, 8, 11, 13, 84, 93. incapacitation, 23, 157, 161–162,
See also general problem solver 166–172, 176
external knowledge, 7–9, 24 incarceration, 162, 164, 167, 169. See
eye for an eye, 157 also imprisonment
indexing, 10
factory robots, 13 indifference, 47–49, 56, 58, 61–63
factual element requirement (of crimi infancy, 28, 123–126
nal liability), 8, 29–32, 35; and gen in personam, 122–123
eral defenses, 118, 127, 136; and in rem, 122, 140, 153
intentional offenses, 38–39, 41–46, insanity, 28, 31, 65, 72, 121, 123,
48, 57, 65, 72–73, 82; and negli 126; as general defense, 128–130,
gence offenses, 85–90, 101; and 154–155
strict liability offenses, 105–111 internal knowledge, 7–9
factual mistake, 72, 121, 123, 133, intoxication, 72, 123, 126, 130–133,
135–137, 154 154–155
factual reality, 8, 10, 31–32, 45, 47; irresistible impulse, 129, 132
index . 243
just deserts, 157–158, 160. See also and sentencing, 156–157; and strict
retribution liability, 109–112, 114–115
justifications, 96, 112, 122–123, 140 monitoring, 4
moral accountability, 17–19, 65
Kant, 66 morality, 18, 26, 65–66, 153
legal mistake, 123, 136–138 necessity, 120, 123, 140, 144–149, 155
legal social control, 26–28 negligence, 32–33, 33, 36–38; and
Leibniz, 1 general defenses, 125, 135, 144;
lesser of two evils, 146–147, 149 and intentional offenses, 45, 68,
loss of self-control, 39, 47, 123, 72, 74, 79–80, 82; as offenses’ type,
126–128, 154–155 84–103; and strict liability, 115,
115–118
machina sapiens, 1, 4–5, 7, 10–11, 13, nullity, 47
17–21, 55, 61, 68
machina sapiens criminalis, 1, 12–14, obedience, 151
17, 19, 21; and intentional offenses, objectivity, 90, 94
55, 61, 68, 82–83; and negligence, omission, 32–33, 36, 39–41, 86–87, 90
101; and strict liability offenses
110–111, 117–118 Pascal, 1
machine learning, 4, 11; in general de perception, 3, 49–52, 102, 131, 133,
fenses, 125, 146, 149, 151; in inten 136
tional offenses, 62, 64, 75; in negli police robots, 13
gence offenses, 84, 92–95, 97, 100; prediction, 10
and sentencing, 160–161, 171–173; prison guard robots, 13, 18
in strict liability offenses, 114, 125, probable consequence liability: and
146, 149, 151 general defenses, 138; in intentional
mechanical computer, 1 offenses, 69, 75–81, 83; in negli
mens rea, 32–33, 33, 36–37; and gen gence, 100–101, 103; in strict liabil
eral defenses, 120–121, 129–131, ity 116–117, 119
135–136, 147; and intentional of probation, 162–163, 165, 169–173
fenses, 44–49, 51–52, 56–57, 63–64, proportionality, 144
67–69, 74, 79–80, 82, and negli prostitution, 13
gence, 85–86, 89–91, 95–96, public service, 165, 167, 171–175
98–102; and strict liability, punishability, 19
105–112, 114–117, 115
mental element requirement (of crimi rashness, 47–48, 49, 56, 58, 61, 63–64
nal liability), 29, 31–41; and general reasonableness, 24, 56, 144
defenses, 118–119, 121, 127–128, recidivism, 167–168
135–137, 139; and intentional of recklessness, 28, 31, 48, 49, 56, 58,
fenses, 44–45, 51, 62, 64–65, 67, 61–63, 82; and negligence, 85–86
69, 72–77, 79–80, 82–83; and negli rehabilitation, 122, 157, 159–162, 166,
gence, 84–87, 89, 95–101, 103–104; 168–172, 176
index . 244
results (factual element component), gence, 85; offenses, criminal liabil

30–32, 31; in intentional offenses, ity for, 104–119, 115; and general
38, 43–52, 49 55, 57–60, 71; in neg defenses, 135, 137
ligence offenses 85, 87–88, 89, 92, substantive immunity, 123, 138, 140
97, 100; in strict liability offenses, superior orders, 123, 149–152, 155
105, 108–109, 109
retribution, 157–160, 166–170, 172, teleological morality, 18, 65
174. See also just deserts thinking machine, 1, 4–5
Turing, 5–7, 14
self-defense, 28, 65, 120–123, 140–142,
144–149, 154–155 ultra vires, 36, 71
semi-innocent agent, 69, 72, 74, 97–99,
114–115, 115 vicarious liability, 35–36, 69–71
sex with robots, 13 volition, 31–32, 33, 45–49, 61–63, 88
slavery, 69
society’s absolute monopoly on power, Watson, 11
140, 144, 147 willful blindness, 52, 55–56
specific intent, 44, 47–48, 57, 59
strict liability, 32–33, 33; and inten zoological legal model, 22–25
tion, 36–38, 45, 68, 74; and negli

When Robots Kill - Artificial Intelligence Under Criminal Law

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

When Robots Kill - Artificial Intelligence Under Criminal Law

Uploaded by

Copyright:

Available Formats

Northeaster n University Press . Boston

University Press of New England is a member of the Green Press

For permission to reproduce any of the material in this book,

Library of Congress Cataloging-in-Publication Data

List of Tables . xiii

3.1. The Factual Element Requirement . 85

4.1. The Factual Element Requirement . 105

4.3. Criminally Liable Entities for

5.1. The Function of General Defenses in Criminal Law . 121

6.1. Conceptual Applicability of Criminal

6.1.2. The Legal Technique of Conversion:

1.1. Target Questions of the Factual Element Components . 31

In 1981, a thirty-seven-year-old Japanese employee in a

In most developed countries, unmanned vehicles, surgical robots, indus-

1. What can be considered as an artificial intelligence entity?

1. Is criminal liability applicable to machines?

The first question is a core issue of criminal law, and imposition of

1.1. The Endless Quest for Machin a Sapiens

1.1.1. H istory and P rehistory of A rtificial I ntelligence

bage, inventor and designer of the analytical engine — the first mechanical,

enthusiastic public, creating at the same time unrealistic expectations and

expert systems resulted in intensive research in the areas of machine learn-

1.1.2. Defining A rtificial I ntelligence, and the E ndlessness

machines that have the ability to think independently would be an impor-

sation between a machine and a human. The conversation is carried out in

selves. This parallels self-­awareness. An intelligent entity is supposed to

Some of these tests have had to do with knowledge representation

1. To possess personhood, an entity must have attribute X.

Therefore, ai technology is not really intelligent. Some scholars called

The contents of attribute X may be made up entirely of advanced in-

and integration of ai entities in industry. The industrial use of ai technol-

1.1.3. Turning Disadvantages into A dvantages:

ful answers to requests of every type. How do these representatives re-

1.2. On Evolu tion and Devolu tion :

1.2.1. Serve and P rotect: H uman F ears of H uman-­R obot C oexistence

and had a small following. Today, Hollywood brings science fiction to

1.2.2. E volution and Devolution:

1. Is criminal liability applicable to machines?

is capable of performance that was considered science fiction only a few

although ai research considers such absence a disadvantage.71 One type of

1.2.3. I napplicability of the Zoological

the robot is a humanoid, and even if the robot is programmed to respond

1.3. The Moder n Offe nder

1.3.1. Basic R equirements for a G iven O ffense

1.3.2. R equirements of a G iven O ffense:

Table 1.1 . Target Questions of the Factual Element Components

f ac tual elemen t componen ts que stion s

The modern structure of the mental element requirement has been

are three principal forms of mental element, which are differentiated

Table 1.2 . Schematic Comparison of the Requirements

Form Cognition Volition

Mens rea subjective cognition only in result offenses

Table 1.3 . Schematic Comparison of the Requirement

mens rea negligence mens rea negligence or

Nevertheless, it is clear that the structure of criminal liability has

1.3.3. The C ase of N onhuman C orporations (R ound 1)

the courts to impose criminal liability on corporations in a way that was

In 1981, a thirty-­seven-­year-­old Japanese employee in a motorcycle factory

2.1. The F ac tual Element R equire ment

Consequently, criminal law considers as an act any material perfor-

accept this modern approach, and there is no need to explicitly require

2.1.3. R esults and C ausal C onnection

1. What can be considered as an artificial intelligence entity?

selves. This parallels self-awareness. An intelligent entity is supposed to

1.2.1. Serve and P rotect: H uman F ears of H uman-R obot C oexistence

In 1981, a thirty-seven-year-old Japanese employee in a motorcycle factory

1. wants the victim to die, it is intent (or specific intent);

Conduct offense awareness of conduct —

2.3.2. H uman L iability: P erpetration-through-A nother

1. a factual condition — the unplanned offense should be the con-

Thirty-thousand victims a year appears to be an enormous number,

weights assigned to some factors. These mistakes are the by-product of

negligence offenses is based on analysis of the machine-learning process

1. Was the a i system unaware of the factual components?

3.3.2. H uman L iability: P erpetration-through-A nother

element requirement of negligence offenses can be considered as semi-